一条视频获得多少播放，取决于它在什么场景下被提供给哪些用户、这些用户可能得到什么价值，以及内容能否持续满足平台的质量和治理要求。观看、点赞、收藏、评论、分享和关注都参与这个过程，但它们所表达的含义并不相同：有人为消遣停留，有人为解决问题收藏，有人为持续学习关注，也有人因为不同意见进入评论区。How many views a video gets depends on the context in which it is presented to users, the potential value those users may gain, and whether the content can continuously meet the platform's quality and governance requirements. Watching, liking, saving, commenting, sharing, and following all participate in this process, but their meanings are not the same: some people stay for entertainment, some save to solve problems, some follow for continuous learning, and some enter the comment section because of differing opinions.
理解新媒体平台，需要同时看三条线：推荐系统怎样筛选内容，价值模型怎样评价用户与创作者的长期收益，治理体系怎样影响内容的传播范围。三者共同组成实际的内容分发系统。Understanding social media platforms requires looking at three lines simultaneously: how the recommendation system filters content, how the value model evaluates the long-term benefits for users and creators, and how the governance system affects the reach of content. Together, these three comprise the actual content distribution system.
本文以参考文献所列平台安全与信任中心的“信息公开”栏目及原始子页面为案例，覆盖截至2026年9月4日的资料。文中具体公式、产品功能、时间线与治理数据均属于这些公开案例，不代表所有新媒体平台采用相同实现。技术解释、案例分析与创作方法按主题展开，参考文献显示名称统一使用“新媒体平台”，原始链接保留于文末供核对。This article uses the Information Disclosure section and original pages of the platform safety and trust center listed in the references as case studies, covering materials available through September 4, 2026. Specific formulas, product features, timelines and governance data belong to these documented cases and do not imply that all social media platforms use the same implementation. Technical explanations, case studies and creator guidance are organized by topic. Reference display names use “Social Media Platform”; the original links are retained at the end for verification.
阅读目录Contents
一、理解推荐系统：它在解决什么问题1. Understanding the Recommendation System: What Problem Does It Solve
1.1 推荐系统的基本任务1.1 Basic Tasks of Recommendation Systems
推荐系统的任务，是在一个远超个人阅读能力的内容集合中，寻找某个用户此刻更可能需要的内容。它面对的不是一份固定节目单，而是持续增长的视频、图文、直播和其他信息供给，以及不断变化的用户行为。The task of a recommendation system is to find the content that a particular user is more likely to need at a given moment from a collection far beyond an individual's reading capability. It is not dealing with a fixed program list, but with a continuously growing supply of videos, texts, live broadcasts, and other information, as well as constantly changing user behavior.
不同的信息组织方式，对用户的主动性要求不同。门户网站先把信息分门别类，用户沿目录寻找；搜索引擎等待用户提出问题，再返回匹配结果；社交网络依赖关注和关系链组织内容；推荐系统则根据历史互动和当前场景，主动推测用户可能需要什么。它们可以共存，搜索和关注行为也可以反过来帮助推荐系统。Different ways of organizing information require different levels of user initiative. Portal websites first categorize information, and users look along the directories; search engines wait for users to ask questions and then return matching results; social networks rely on followers and relationship chains to organize content; recommendation systems proactively infer what users may need based on historical interactions and the current context. They can coexist, and searching and following behaviors can in turn assist the recommendation system.
案例平台把推荐算法描述为高效的个性化信息过滤工具。一个用户曾经看过什么、怎样观看、对哪些内容表达认可、在什么情况下快速离开，都会构成理解需求的线索。内容本身的特征，以及当前时间等场景信息，也参与匹配。Case platforms describe recommendation algorithms as efficient personalized information filtering tools. Clues for understanding user needs include what a user has watched before, how they watched it, which content they expressed approval for, and the circumstances under which they quickly left. Features of the content itself, as well as contextual information like the current time, also participate in matching.
因此，一条视频的推荐价值与接收它的人有关。专业相机维修视频可能对摄影器材爱好者很有价值，对没有相关需求的用户吸引力很低。推荐系统试图解决的正是这种差异：让同一条内容在合适的人群中发挥价值，而不是简单给所有人播放同一批高热视频。Therefore, the recommendation value of a video depends on the person receiving it. A professional camera repair video may be very valuable to photography equipment enthusiasts but of little interest to users without relevant needs. The recommendation system aims to address this difference: allowing the same content to deliver value to the appropriate audience, rather than simply playing the same batch of high-quality videos to everyone.
1.2 推荐、热门和热榜是三个不同概念1.2 Recommendation, Popularity, and Hot Lists Are Three Different Concepts
推荐，是内容被系统选入某个用户的信息流。热门，是人们对内容获得较大传播量的日常描述。热榜，则是特定产品中的榜单结果，通常对应明确的榜单对象、排序和运营机制。一个作品被大量个性化推荐，与某个话题进入热榜，并不是同一件事。Recommendation is when content is selected by the system into a particular user's feed. Popularity is a general description of content gaining a large amount of spread among people. Hot lists are the ranking results in specific products, usually corresponding to clear ranking targets, order, and operational mechanisms. A piece of content being widely recommended in a personalized manner is not the same as a topic entering a hot list.
讨论“为什么这条视频火了”时，可以进一步拆成四个问题：它为什么进入候选集合，为什么在部分用户那里排得靠前，为什么能够不断触达新的相关用户，以及为什么在传播扩大后仍然适合继续推荐。前两个问题主要涉及匹配和排序，第三个涉及需求规模、反馈和探索，第四个同时涉及质量与治理。When discussing "why this video went viral," it can be further broken down into four questions: why it entered the candidate set, why it ranks high for certain users, why it can continuously reach new relevant users, and why it is still suitable for continued recommendation after its spread expands. The first two questions mainly involve matching and ranking, the third involves demand scale, feedback, and exploration, and the fourth involves both quality and governance.
这样的拆分有助于理解播放曲线。一条视频停止增长，可能是相关用户被逐步覆盖，也可能是后续受众的反应发生变化，或是内容进入了新的治理评估。单凭一个播放量数字，无法分辨这些环节各自发挥了多大作用。This breakdown helps to understand the view curve. A video's growth may stop because relevant users have been gradually covered, because the reactions of subsequent audiences change, or because the content enters a new governance evaluation. A single view count number cannot distinguish how much each of these links has contributed.
本文所说的热门机制，主要讨论个性化推荐如何形成大规模传播。栏目中的公开资料提供了这条链路的原理、价值目标和治理节点，没有给出一份可直接计算所有视频最终播放量的统一排行榜公式。The popular mechanism discussed in this article mainly addresses how personalized recommendations generate large-scale dissemination. The publicly available information in the section provides the principles, value goals, and governance nodes of this chain, but does not provide a unified formula that can directly calculate the final view count for all videos.
1.3 行为预测、价值判断与内容理解1.3 Behavior Prediction, Value Judgment, and Content Understanding
推荐系统里有几类容易混淆的能力。行为预测回答用户是否可能看完、点赞、收藏或关注；价值判断回答不同反馈应该怎样影响推荐；内容理解则尝试识别画面、声音、语言、语境及它们之间的关系。这些能力可以相互配合，但并不等价。There are several easily confused capabilities in recommendation systems. Behavior prediction answers whether a user is likely to finish watching, like, favorite, or follow; value judgment addresses how different feedback should influence recommendations; content understanding tries to recognize images, sounds, language, context, and the relationships among them. These capabilities can complement each other but are not equivalent.
例如，一段强烈刺激情绪的视频可能让人停留很久。模型预测到停留，并不自动意味着内容真实、对用户长期有益，或适合在整个社区扩大传播。平台因此需要多目标价值模型和治理流程，补充单个即时行为无法表达的信息。For example, a video with highly stimulating emotions may hold attention for a long time. Model prediction of attention does not automatically mean the content is truthful, beneficial to users in the long term, or suitable for broad dissemination throughout the community. Therefore, the platform needs multi-objective value models and governance processes to supplement information that a single instant behavior cannot convey.
2025年初的基础科普着重说明：即使不具备人的理解方式，模型也能通过用户与内容的统计关联预测行为。2025年末的大模型文章进一步讨论语义理解、上下文建模与知识推理的作用。两类文章对应不同层次的说明：前者解释推荐为何可以依靠数学关系成立，后者讨论如何使用新的模型能力增强推荐。The basic science popularization at the beginning of 2025 emphasized: even without human understanding, models can predict behavior through statistical associations between users and content. The large-model articles at the end of 2025 further discuss the roles of semantic understanding, context modeling, and knowledge reasoning. These two types of articles correspond to different levels of explanation: the former explains why recommendations can rely on mathematical relationships, while the latter discusses how to use new model capabilities to enhance recommendations.
判断推荐系统“懂不懂内容”，更准确的方式是看具体任务。识别字幕、理解反讽、推测用户需求、查证事实和进行价值判断，各有不同的输入、训练方式与评估目标。把它们合成一句“算法懂”或“算法不懂”，反而会遮蔽实际机制。A more accurate way to judge whether a recommendation system "understands the content" is by looking at specific tasks. Recognizing subtitles, understanding irony, speculating user needs, verifying facts, and making value judgments each have different inputs, training methods, and evaluation goals. Combining them into a sentence of "the algorithm understands" or "the algorithm doesn't understand" can actually obscure the actual mechanism.
二、一条视频怎样进入用户的信息流2. How a Video Enters a User's Feed
2.1 从内容池到最终列表2.1 From the Content Pool to the Final List
传统推荐流程可以理解为逐层收缩候选范围：内容发布并接受初步评估，进入可供推荐的内容集合；召回从海量内容中筛出一批相关候选；排序进一步判断哪些候选更值得优先展示；最后再通过多样性、新鲜度和产品策略调整列表。The traditional recommendation process can be understood as narrowing down the candidate pool step by step: content is published and initially evaluated, entering the collection of recommended content; Recall filters out a batch of relevant candidates from a vast amount of content; Sorting further determines which candidates are more worth prioritizing; Finally, the list is adjusted through diversity, freshness, and product strategy.
公开文章把更完整的传统链路描述为召回、粗排、精排、混排等环节。召回重视覆盖与效率，粗排使用相对轻量的计算快速缩小范围，精排投入更复杂的模型估计多个目标，混排或重排则处理最终列表的整体体验。某条视频单独看起来不错，也可能因为列表里已经有相似内容而被调整位置。Public articles describe the more complete traditional chain as recall, rough queue, fine queue, mixed queue, and so on. Recall emphasizes coverage and efficiency; rough queue uses relatively lightweight calculations to quickly narrow the scope, fine queue invests more complex models to estimate multiple targets, and mixed or rearranged processes the overall experience of the final list. A video that looks good on its own may be repositioned because similar content already exists in the list.
互动科普用“从大量内容中找出数千条候选，再筛选出更小集合”的方式展示这一过程。这里的数量是帮助理解量级的说明。实际服务会受产品场景、模型配置、内容供给和用户情况影响。Interactive science popularization demonstrates this process by "identifying thousands of candidates from a large amount of content and then filtering for smaller collections." The quantity here is an explanation to help understand the scale of magnitude. Actual services are affected by product scenarios, model configuration, content supply, and user situations.
治理与这条漏斗同时运行。初始发布时检查底线问题，传播过程中仍可能重新评估，因此内容进入候选池不等于此后永久保有推荐资格。匹配、排序、探索和治理并不是互相隔绝的部门流程，而是共同作用于最终分发结果的系统环节。Governance runs simultaneously with this funnel. Bottom-line issues are checked at initial release, and may still be re-evaluated during dissemination, so content entering the candidate pool does not mean permanent recommendation eligibility. Matching, sequencing, exploration, and governance are not isolated departmental processes, but rather systematic links that jointly influence the final distribution results.
2.2 为什么需要多条召回通路2.2 Why Multiple Recall Routes Are Needed
如果只用一种方式找候选内容，系统容易漏掉不同来源的需求。用户可能同时对历史、运动和家庭烹饪感兴趣，也可能在今天突然需要某个维修教程。召回因此需要从多个方向寻找内容，随后合并候选集合。If you only find candidate content in one way, the system can easily miss the needs of different sources. Users may be interested in history, sports, and home cooking all at once, or suddenly need a repair tutorial today. Recalls therefore require searching for content from multiple directions and then merging candidate collections.
基础文章介绍双塔召回，多样性文章介绍多兴趣召回、长尾扶持与随机探索，新版兴趣模型则把长期统计引入召回。科普课程的简介还用热门与兴趣等不同“大网”解释候选来源。不同通路可以承担不同任务：发现已经表现出的兴趣、补足被低估的兴趣、恢复长期偏好，或试探尚未出现明确反馈的新内容。The basic article introduces dual-tower recall, diverse articles introduce multi-interest recall, long-tail support, and random exploration, while the new interest model introduces long-term statistics into recall. The introduction of popular science courses also explains the sources of candidates using different “macro nets” such as popularity and interests. Different pathways can undertake different tasks: discovering already expressed interests, supplementing underestimated interests, restoring long-term preferences, or probing new content that has not yet received clear feedback.
召回的意义在于争取被后续模型评估的机会。某条内容即使潜在价值很高，如果没有进入候选集合，后续精排也无法把它排到前面。这也是新一代架构探索希望减少多层信息损失的原因。The significance of recall lies in securing the opportunity to be evaluated by subsequent models. Even if a piece of content has high potential value, if it does not enter the candidate set, the subsequent ranking cannot place it at the forefront. This is also the reason why the next-generation architecture aims to reduce multi-layer information loss.
对新视频而言，系统最初拥有的互动记录有限，需要结合已有特征与探索逐步积累反馈。对新用户而言，也需要在有限历史下试探需求。公开科普把这些问题归入冷启动和探索，但没有给出所有新作品都一致的首批播放配额。For new videos, the system initially has limited interaction records and needs to gradually accumulate feedback by combining existing features with exploration. For new users, it is also necessary to probe their needs based on limited history. Public science explanations categorize these issues as cold start and exploration but do not provide a uniform first batch of play quotas for all new works.
2.3 双塔模型如何完成匹配2.3 How the Dual-Tower Model Completes Matching
双塔模型分别处理用户侧与内容侧的信息。用户塔接收历史行为、兴趣和场景等特征，内容塔接收视频或物品的相关特征。经过训练，两端输出可以在同一数学空间中比较的向量。The dual-tower model processes user-side and content-side information separately. The user tower receives features such as historical behaviors, interests, and scenarios, while the content tower receives relevant features of videos or items. After training, the outputs from both sides are vectors that can be compared within the same mathematical space.
向量可以理解为一串数字组成的表示。现实中的多种因素被压缩到这个表示里，使系统能够高效比较大量用户与内容之间的关系。训练后的每一维通常不再对应一个容易直接命名的标签，因此不能把向量中的某个数字简单理解为“喜欢美食程度”或“爱看搞笑程度”。Vectors can be understood as representations made up of a string of numbers. Various real-world factors are compressed into this representation, allowing the system to efficiently compare relationships between a large number of users and content. Each dimension after training usually no longer corresponds to an easily directly nameable label, so a number in the vector cannot simply be interpreted as “degree of liking food” or “degree of enjoying comedy.”
在召回时，系统利用相似度或匹配分寻找候选。用户向量与某个内容向量越匹配，该内容越可能进入候选集合。具体匹配函数取决于模型设计；向量接近表达的是模型中的相关性，不直接等于最终观看概率，更不等于热门概率。During recall, the system uses similarity or matching scores to find candidates. The closer the user vector matches a content vector, the more likely the content is to enter the candidate set. The specific matching function depends on the model design; vector closeness expresses the correlation in the model, which does not directly equal the final probability of watching, and even less, the popularity probability.
双塔结构适合大规模匹配，是因为内容端表示可以被高效组织和检索。它解决的是从巨大候选空间中迅速找到相关内容的问题，后续排序还要结合更多目标和信息进一步判断。The dual-tower structure is suitable for large-scale matching because content-side representations can be efficiently organized and retrieved. It addresses the problem of quickly finding relevant content from an enormous candidate space, while subsequent ranking still needs to combine more objectives and information for further judgment.
一个直观例子是，长期观看相机参数测评的人，可能也会对镜头保养、摄影包选择和后期工作流感兴趣。模型可以从交互模式中学到这种联系，而不必要求用户先手工填写一张完整的兴趣调查表。A straightforward example is that people who watch long-term camera parameter reviews may also be interested in lens maintenance, choosing camera bags, and post-production workflows. The model can learn this connection from interaction patterns without requiring the user to manually fill out a complete interest survey.
2.4 相似度公式怎样计算2.4 How is the similarity formula calculated
原网页用余弦相似度介绍向量关系。设用户表示为向量 (u)，内容表示为向量 (v)，余弦相似度为：The original webpage introduces the vector relationship using cosine similarity. Let the user be represented as vector (u) and the content as vector (v). Cosine similarity is:
分子是逐维相乘再求和，分母是两个向量长度的乘积。这个指标主要比较方向。非零向量同向时结果接近1，互相垂直时为0，方向相反时接近-1。The numerator is the sum of the products of corresponding dimensions, and the denominator is the product of the lengths of the two vectors. This metric mainly compares direction. Non-zero vectors pointing in the same direction have a result close to 1; if perpendicular to each other, it is 0; and if pointing in opposite directions, it is close to -1.
例如 (u=(1,2))、(v=(2,3))：For example, (u=(1,2)), (v=(2,3)):
结果表明两个表示在这个简化空间里方向接近。0.9923是相似度值，而不是99.23%的点赞概率或热门概率。若需要得到行为概率，通常还要通过进一步的模型计算和概率映射。The result indicates that the two representations are close in direction in this simplified space. 0.9923 is the similarity value, not a 99.23% like probability or popularity probability. To obtain behavior probabilities, further model computation and probability mapping are usually required.
余弦相似度、点积和距离是不同的比较方式。向量归一化以后，余弦相似度与点积会建立直接联系；未归一化时，向量长度也会影响点积。因此，不能仅凭科普文章展示一种公式，就推断所有召回模型使用完全相同的打分方式。Cosine similarity, dot product, and distance are different comparison methods. After vector normalization, cosine similarity and dot product have a direct relationship; without normalization, vector length also affects the dot product. Therefore, one cannot infer that all recall models use exactly the same scoring method merely from a popular science article presenting a formula.
2.5 兴趣时钟：同一个人为什么在不同时段看不同内容2.5 Interest Clock: Why the same person views different content at different times
推荐不仅涉及用户与内容，还涉及场景。一个人早晨通勤可能偏好简短资讯，周末下午可能愿意看出游攻略，晚上有较长空闲时才阅读深入分析。这些是兴趣与情境之间的关系，而不是用户突然变成了另一类人。Recommendations involve not only the user and content but also the context. A person commuting in the morning may prefer short news, on a weekend afternoon might be willing to read travel guides, and only in the evening with a longer free period engages with in-depth analysis. These relationships reflect the interaction between interest and context, not that the user has suddenly become a different type of person.
“体验算法”用兴趣时钟展示时间因素如何影响推荐倾向。用户拨动时针，可以看到演示中的内容组合改变。它说明时间可以成为模型输入，帮助区分同一用户在不同场景下的需要。The "experience algorithm" uses the interest clock to demonstrate how time factors influence recommendation tendencies. By moving the clock hands, users can see content combinations change in the demo. It shows that time can be an input for the model to help distinguish the same user's needs in different contexts.
新版推荐文章同样讨论了历史偏好与实时场景的联合建模。例如，模型可以结合某人对露营的长期兴趣、近期的装备浏览，以及当前是否接近周末，估计其对某类内容的需求。The updated recommended articles also discuss joint modeling of historical preferences and real-time contexts. For example, the model can combine someone's long-term interest in camping, recent browsing of gear, and whether the weekend is near to estimate their demand for a type of content.
从这个机制可以推导出一个创作实验方向：观察内容所解决的问题通常在何时出现，再比较相近作品在相近场景下的反馈。早餐教程、通勤信息和周末活动指南可以有不同的需求节奏。场景匹配的重点是让内容在需求出现时有用，公开资料没有把某个时刻定义为适用于所有账号的统一发布黄金时间。From this mechanism, one can derive a direction for creative experimentation: observe when the problems addressed by the content typically arise, and then compare the feedback of similar works in similar scenarios. Breakfast tutorials, commuting information, and weekend activity guides may have different demand rhythms. The key to scenario matching is to make the content useful when the demand arises; public information does not define a unified 'golden time' for all accounts.
三、推荐分数与多目标公式3. Recommendation Scores and Multi-Objective Formula
3.1 概率模型与价值模型的分工3.1 The Division of Labor between Probability Models and Value Models
推荐排序需要先估计用户可能产生哪些行为，再评价这些行为对用户、作者和平台意味着什么。前者是概率模型的任务，后者是价值模型的任务。Recommendation ranking first needs to estimate what actions users might take, and then evaluate what these actions mean for users, creators, and the platform. The former is the task of the probability model, and the latter is the task of the value model.
原网页把核心逻辑概括为：综合预测用户行为概率，与行为价值权重结合，形成视频推荐优先级。配图列出了用户点赞概率、收藏概率、评论概率等一组行为预测，以及用户价值、长期用户价值、作者价值、平台价值等一组价值维度。The original webpage summarizes the core logic as: combining the predicted probabilities of user actions with the value weights of those actions to form the video recommendation priority. The accompanying image lists a set of behavior predictions such as the probability of likes, collections, and comments, alongside a set of value dimensions including user value, long-term user value, creator value, and platform value.
可以将其理解为两个连续问题：第一，如果把视频推荐给这个用户，他可能怎样反应；第二，这些反应在当前推荐目标下有多重要。一条视频可能在某种行为上占优、在另一种行为上不占优，综合排序需要处理这种差异。This can be understood as two consecutive questions: first, if this video is recommended to this user, how might they react; second, how important are these reactions under the current recommendation objectives. A video may excel in one type of behavior but not in another, and the comprehensive ranking needs to handle this difference.
配图还出现了符号 (\delta)，但页面正文未给出它的完整定义。可确定的信息是概率与价值共同参与推荐，而不是把某个未解释符号自行解释为账号等级、流量池系数或人工加权系数。The image also shows the symbol (\delta), but the main text does not provide its full definition. What can be determined is that probability and value jointly participate in recommendation, rather than interpreting an unexplained symbol as account level, traffic pool coefficient, or manual weighting factor.
“行为价值”也不是用户已经点了多少赞的简单累计。排序发生在展示之前，模型需要估计一个尚未发生的用户—视频交互。历史真实反馈用于帮助模型学习，预测值用于当前候选的比较，两者在时间和用途上不同。'Behavior value' is also not a simple accumulation of how many likes the user has already given. The ranking occurs before display, and the model needs to estimate a user-video interaction that has not yet happened. Historical real feedback helps the model learn, while predicted values are used for comparing current candidates; the two differ in time and purpose.
3.2 原网页中的多目标乘法表达3.2 The Multi-Objective Multiplicative Expression on the Original Webpage
参考文献 S03 所列多目标平衡文章的配图展示了如下简化形式：The figure in Reference S03 on multi-objective balancing shows the following simplified form:
其中，(S)表示推荐分数；like、follow、comment分别对应点赞、关注与评论目标；(value)和(weight)表示目标价值及相关权重；(f)表示对某个目标进行计算或转换的函数。省略号意味着还包括其他目标。Here, (S) represents the recommended score; like, follow, and comment correspond to the targets of liking, following, and commenting, respectively; (value) and (weight) represent the target value and its associated weight; (f) represents a function that calculates or transforms a certain target. The ellipsis implies that other targets are also included.
这个表达最重要的信息，是不同目标经过处理后共同决定分数。图中采用乘法形式，不能把它转写成“点赞数加评论数”后当成同一个公式。与此同时，函数 (f) 的具体形式、输入如何标定、权重如何取值以及不同场景如何组合，都没有在这张科普图里展开。The most important information conveyed by this expression is that different targets collectively determine the score after processing. The diagram uses a multiplicative form, which cannot be rewritten as "number of likes plus number of comments" and treated as the same formula. Meanwhile, the specific form of the function (f), how the input is calibrated, how the weights are set, and how different scenarios are combined are not elaborated in this popular science illustration.
从数学上看，乘法容易表达目标之间的共同作用：多个大于基准的因素相乘，会产生叠加放大；某个因素降低，也可能抑制整体结果。但实际系统需要处理零值、数值范围、稳定性和目标间尺度差异，因此不能假定“没有评论就一定总分为零”。关键在于函数如何把原始信号变换成参与融合的数值。Mathematically, multiplication easily expresses the combined effect of targets: multiple factors above the baseline multiply to produce cumulative amplification; if one factor decreases, it may suppress the overall result. But in an actual system, zero values, value ranges, stability, and scale differences among targets need to be addressed, so one cannot assume "no comments automatically means a total score of zero." The key lies in how the function transforms the raw signals into values used in the fusion.
这也是理解公式时必须保留的结构：目标值、目标权重、目标变换和最终融合，是四个相关但不同的部分。This structure must also be retained when understanding the formula: target values, target weights, target transformation, and final fusion are four related but distinct parts.
3.3 用计算示例理解权重怎样改变结果3.3 Understanding how weights change results through calculation examples
为了说明概率与权重的关系，可以先构造一个教学用的加权评分：To illustrate the relationship between probability and weight, we can first construct a weighted scoring example for teaching purposes:
假设视频A的点赞、收藏、关注概率分别为0.20、0.03、0.01，视频B分别为0.12、0.08、0.03。计算得到：Suppose video A has predicted probabilities for like, favorite, and follow of 0.20, 0.03, and 0.01, respectively, and video B has 0.12, 0.08, and 0.03. Calculation yields:
视频A的预测点赞概率更高，但B在收藏和关注上更强，因此在这组假设权重下得分更高。这个例子解释了为什么单个指标更高，不足以决定综合排序。这里的加权和是解释工具，不是对原网页乘法图的替换。Video A has a higher predicted like probability, but B is stronger in favorite and follow, so under this set of assumed weights, B scores higher. This example explains why a single higher metric is not sufficient to determine the overall ranking. The weighted sum here is a tool for explanation, not a replacement for the original multiplicative diagram on the webpage.
再用一个乘法教学例子：令 (f_i=1+w_i p_i)。则A的得分为 (1.20\times1.12\times1.06\approx1.4246)，B为 (1.12\times1.32\times1.18\approx1.7445)。它展示了先变换、再融合的过程，并且在某个概率为零时，该因子仍可取基准值1。Another example of multiplication teaching: let (f_i = 1 + w_i p_i). Then A's score is (1.20 × 1.12 × 1.06 ≈ 1.4246), and B's score is (1.12 × 1.32 × 1.18 ≈ 1.7445). It demonstrates the process of transforming first, then merging, and even when a certain probability is zero, the factor can still take the baseline value of 1.
两个例子中的权重和函数均为演算设定。它们的用途是帮助理解多目标关系：不同融合形式可能产生不同排序，因此研究公开机制时，应把公式结构与未公开参数分别对待。In both examples, the weights and functions are set for calculation purposes. Their use is to help understand multi-objective relationships: different forms of fusion may produce different rankings, so when studying public mechanisms, the formula structure and unpublished parameters should be treated separately.
3.4 推荐系统会预估哪些行为3.4 Behaviors that the recommendation system predicts
公开正文与配图涉及的行为大致可以分为四组。第一组是基础观看，包括观看时长、是否看完、是否较快划走等；第二组是显性交互，包括点赞、评论、收藏、分享、关注、点击作者头像以及对内容表达不喜欢；第三组是长期消费，包括后续再次打开、主动搜索、持续关注和回访；第四组涉及内容供给和创作者体验，例如投稿及互动反馈。The behaviors involved in the main text and illustrations can roughly be divided into four groups. The first group is basic viewing, including watch time, whether the video was finished, whether it was scrolled quickly, etc.; the second group is explicit interaction, including likes, comments, saves, shares, follows, clicking on the author’s avatar, and expressing dislike for the content; the third group is long-term consumption, including reopening, actively searching, continuously following, and revisiting; the fourth group involves content supply and creator experience, such as submissions and interaction feedback.
评论区停留时长也被列为预估行为之一。这意味着观看不只发生在视频画面上，用户围绕内容阅读讨论的行为也可能提供信息。不同信号共同描绘用户体验，不能只以视频播放结束与否概括全部消费过程。Time spent in the comment section is also listed as one of the predicted behaviors. This means that viewing does not only occur on the video screen; users’ behavior in reading and discussing content may also provide information. Different signals together depict the user experience and cannot be summarized solely by whether the video playback ends.
行为信号的含义需要结合语境。点赞可能是认可，也可能是礼貌；收藏可能是反复使用，也可能是暂时没时间看；关注可能代表对一个人的持续兴趣，也可能是等待某个后续结果。组合目标和长期观察正是为了进一步分辨这些差异。The meaning of behavioral signals needs to be considered in context. A like may indicate approval or politeness; saving may indicate repeated use or temporary unavailability to watch; following may represent sustained interest in a person or waiting for a future outcome. Combining objectives and long-term observation is precisely to further distinguish these differences.
因此，多目标系统关注的不只是动作数量，还包括动作发生在什么用户、什么内容、什么场景中，以及后续是否延续成更持久的关系。对创作者而言，理解这套体系，比把所有互动统一叫作“热度”更有解释力。Therefore, a multi-objective system focuses not only on the quantity of actions but also on which users, which content, and which scenarios the actions occur in, as well as whether they continue into more lasting relationships. For creators, understanding this system is more explanatory than simply labeling all interactions as 'popularity.'
3.5 反馈为什么能快速影响推荐3.5 Why feedback can quickly influence recommendations
用户行为发生后，会成为新的学习信号。推荐系统持续接收这些信号，更新对用户与内容关系的判断。公开文章描述案例平台已经实现分钟级的实时反馈更新，使行为与模型学习之间的间隔缩短。After user behavior occurs, it becomes a new learning signal. The recommendation system continuously receives these signals to update its judgment of the relationship between users and content. Public articles describe that the case platform has already achieved minute-level real-time feedback updates, shortening the interval between behavior and model learning.
这并不意味着某次点赞会被机械地兑换成下一分钟固定数量的曝光。实时更新改变的是模型对相关行为和价值的估计，最终结果仍取决于候选竞争、用户差异和其他目标。This does not mean that a single like will be mechanically converted into a fixed number of exposures in the next minute. Real-time updates change the model's estimation of the relevance and value of the behavior, and the final outcome still depends on candidate competition, user differences, and other objectives.
还需要区分模型参数与价值权重。模型参数用于学习复杂的预测关系，价值权重表达不同目标的相对重要性。两者都可能调整，但不能把“模型实时接收反馈”自动解释成“所有平台价值权重每分钟重设一次”。It is also necessary to distinguish between model parameters and value weights. Model parameters are used to learn complex predictive relationships, while value weights express the relative importance of different objectives. Both can be adjusted, but "real-time feedback reception by the model" should not automatically be interpreted as "all platform value weights being reset every minute."
从内容分析角度看，传播是动态过程。新一批受众带来的反馈可能与早期受众不同，模型会在反馈积累中修正判断。这能解释某些作品开始缓慢、后来找到受众，也能解释某些作品初期互动较好、扩大受众后表现下降。具体变化仍需要结合实际数据，而不是只依据发布时间推断。From a content analysis perspective, propagation is a dynamic process. Feedback from a new batch of audiences may differ from early audiences, and the model will adjust its judgment as the feedback accumulates. This can explain why some works start slowly but later find their audience, and also why some works have good early interaction but perform worse as the audience expands. Specific changes still need to be based on actual data rather than inferred solely from the publishing time.
四、优质内容为什么可能获得更多分发4. Why High-Quality Content May Receive More Distribution
4.1 优质内容子链路与独立分发通道4.1 Sub-path for high-quality content and independent distribution channels
平台公开资料明确提到，在原有推荐链路中增加了面向优质内容的子链路。用于发现优质内容的信号包括收藏后的复访、作者对用户的长期陪伴，以及社区的优质交流氛围。案例平台的精选 App 和精选频道也承担了优质内容承载与消费的作用。The platform's public information clearly mentions that a sub-path for high-quality content has been added in the original recommendation chain. Signals used to discover high-quality content include revisits after collection, long-term engagement of users with the author, and a high-quality community interaction atmosphere. The platform's curated apps and curated channels also play a role in carrying and consuming high-quality content.
2026年的平台举措继续提出构建优质内容池、优化推荐机制、扩大优质创作者的收益，并通过青年创作者、创作阶梯、AI创作浪潮和媒体精选等计划支持不同阶段的内容生产。2026年8月的原网页再次提到优质内容独立分发通道，并将其与懂行用户识别、长尾兴趣模型联系起来。The platform's initiatives in 2026 continue to propose building a pool of high-quality content, optimizing recommendation mechanisms, increasing the income of high-quality creators, and supporting content production at different stages through plans for young creators, creation ladders, AI creation waves, and media curation. The original web page from August 2026 again mentions the independent distribution channels for high-quality content and links them with discerning user identification and long-tail interest models.
独立分发通道的意义，是为某些在即时指标中不一定占优、但具有更高深度或长期价值的内容提供额外被发现的路径。长知识视频、专业讨论和小众文化内容，不必只依赖与轻量娱乐内容完全相同的表面信号进行竞争。The significance of independent distribution channels is to provide an additional path for content that may not necessarily excel in immediate metrics but has higher depth or long-term value. Long knowledge videos, professional discussions, and niche cultural content do not have to rely solely on the same superficial signals used for light entertainment content to compete.
这里的“优质”涉及内容价值及用户体验，而不是单看画质、设备和剪辑成本。清晰表达、原创信息、具体经验、专业认可和持续使用价值，都可能帮助一条作品形成更丰富的正向反馈。公开页面提供了这些机制和案例，但并未公布一张统一的优质入池打分表。Here, "premium" refers to content value and user experience, not just image quality, equipment, or editing costs. Clear expression, original information, concrete experience, professional recognition, and ongoing value can all help a work generate richer positive feedback. Public pages provide these mechanisms and case studies, but do not publish a unified quality pooling score.
4.2 懂行用户怎样帮助识别内容质量4.2 How knowledgeable users can help identify content quality
对于专业内容，所有反馈并不具有完全相同的解释力。看热闹的人可以判断是否有趣，长期接触某个领域的人则更有条件判断信息是否准确、细节是否扎实、方法是否有用。公开资料介绍了“懂行用户”机制：利用这类用户在相关领域的评论、收藏等深度互动，帮助算法发现优质内容。For professional content, not all feedback has exactly the same explanatory power. Onlookers can judge whether it's interesting, while those who have long been involved in a particular field are better positioned to judge whether the information is accurate, the details are solid, and the method is effective. Public information introduces the "expert user" mechanism: using these users to deeply interact with comments and favorites in related fields to help algorithms discover high-quality content.
2026年发布的案例文章，用知识科普创作者说明这种机制。2026年5月的评论治理文章进一步指出，在案例平台精选内容的分发逻辑中，专业领域用户的深度好评论可以作为质量认证信号，触发额外推荐。Case articles published in 2026 use knowledge popularization creators to illustrate this mechanism. The May 2026 review governance article further pointed out that in the distribution logic of selected content on case platforms, in-depth good comments from users in professional fields can serve as quality certification signals and trigger additional recommendations.
这提供了一种理解专业内容传播的方式。比如一条家具修复视频，如果评论区出现同行对工艺细节的讨论、使用者对实际效果的反馈、初学者提出的具体问题，它承载的信息显然比一排无关表情丰富。对内容质量的识别，能够利用这些不同层次的反馈。This provides a way to understand the dissemination of professional content. For example, in a furniture restoration video, if the comments section includes discussions among peers about craftsmanship details, user feedback on actual results, or specific questions from beginners, the information it carries is clearly richer than a row of irrelevant expressions. Recognizing content quality can leverage feedback at these different levels.
“懂行”在这里是与领域有关的判断能力，不应直接等同于粉丝量、职业称谓或账号认证标志。公开页面没有提供完整的识别条件。对创作而言，可操作的方向是提供经得起专业讨论的内容，并让真实使用者和相关领域用户有理由展开交流。"Knowledgeable" here refers to judgment related to the field and should not be directly equated with follower count, professional titles, or account verification marks. Public pages do not provide complete identification criteria. For creative work, the practical direction is to provide content that can withstand professional discussion and give real users and users in related fields a reason to communicate.
4.3 收藏与复访如何识别长期价值4.3 How to Identify Long-Term Value Through Collection and Revisit
收藏是一种比即时停留更明确的保留动作。一个用户收藏某条视频，往往意味着希望以后再次访问，或者认为其中的信息值得保存。平台将收藏入口放到更显眼的位置，并调整相关推荐逻辑，使这类需求更容易形成可学习的信号。Favorites are a more explicit retention action than instant pause. When a user saves a video, it often means they want to visit it again later or believe the information is worth keeping. Platforms place the favorites entry point in a more prominent position and adjust the relevant recommendation logic to make it easier to form learnable signals for these needs.
进一步看，收藏本身仍包含多种情况。用户可能只是正在忙，晚些时候补看；也可能把教程保存下来，在未来几天或更长时间内反复使用。因此，多目标体系会关注收藏与后续复访的组合，而不只观察一次收藏动作。Looking further, collections themselves still encompass various situations. Users might just be busy and plan to view it later; they may also save tutorials to repeatedly use them over the next few days or longer. Therefore, a multi-goal system would focus on the combination of collections and subsequent revisits, rather than just observing a single collection action.
这类组合尤其适合识别操作指南、健身动作、烹饪步骤、学习资料、工具说明和需要持续参考的知识内容。反复访问说明作品进入了用户的实际活动，而不仅是在第一次浏览时获得注意。Such combinations are particularly suitable for identifying operation guides, fitness movements, cooking steps, study materials, tool instructions, and knowledge content that requires ongoing reference. Repeated visits indicate that the work has entered the user's actual activities, rather than just gaining attention during the first browsing.
从内容设计上，可以把“方便回来用”作为一个独立目标：步骤有顺序，关键参数容易找到，必要条件交代清楚，结果和常见错误有对应说明。这样即使用户忘记完整叙述，也能在再次打开时迅速定位所需信息。From a content design perspective, "convenient for later use" can be treated as an independent goal: steps are in order, key parameters are easy to find, necessary conditions are clearly explained, and results and common mistakes have corresponding explanations. In this way, even if users forget the complete narrative, they can quickly locate the needed information when they open it again.
这个推导不是要求所有视频都做成教程。故事、音乐和情感表达也可能被重复消费，只是它们的复访原因不同。核心问题是作品能否在第一次观看之后继续创造价值。This deduction does not mean that all videos should be made into tutorials. Stories, music, and emotional expressions may also be consumed repeatedly, but the reasons for revisiting them are different. The core question is whether the work can continue to create value after the first viewing.
4.4 关注与追更怎样连接作品和作者4.4 How Following and Subscribing Connect Works and Authors
点赞主要围绕当前作品，关注则把关系延伸到作者。用户希望以后继续看到某个人的创作，意味着推荐系统可以尝试理解一种跨作品的持续需求。Likes mainly revolve around the current work, while following extends the relationship to the author. Users wanting to continue seeing someone's creations in the future implies that the recommendation system can attempt to understand a sustained need across works.
关注之后是否追更，有助于区分短暂好奇与稳定关系。某人关注一个做历史讲解的作者，随后持续观看不同主题的视频，说明吸引力可能不仅来自单个选题，还来自作者的讲述方式、知识结构或可靠程度。Whether users subscribe after following helps distinguish between temporary curiosity and stable relationships. If someone follows an author who explains history and then continues watching videos on different topics, it indicates that the attraction may come not only from a single topic but also from the author's presentation style, knowledge structure, or reliability.
公开的多目标体系把关注与追更作为长期组合目标，并讨论作者长期陪伴所形成的价值。创作者的稳定表达、持续供给与用户的反复访问，可以共同构成一种长期关系。Public multi-goal systems treat following and subscribing as long-term combined goals and discuss the value formed by the author's long-term presence. The creator's consistent expression, continuous supply, and the user's repeated visits can jointly constitute a long-term relationship.
系列内容的设计可以围绕同一受众的连续问题展开，而不是把一段本可以完整讲清的内容强行拆开。每一集提供独立价值，系列整体又有明确的递进关系，用户才有真实的追更理由。例如，相机入门可以从曝光理解、镜头选择、实际拍摄延伸到后期处理，而不是每集只在结尾承诺下一集才揭晓答案。Series content can be designed around the continuous questions of the same audience, rather than forcibly splitting content that could be fully explained in one go. Each episode provides independent value, while the series overall has a clear progression, giving users a real reason to subscribe. For example, beginner photography could extend from understanding exposure, lens selection, actual shooting, to post-processing, rather than each episode only promising that the answer will be revealed in the next episode.
对于用户来说，持续关注的是一个可靠的信息来源或熟悉的表达者；对于推荐系统来说，这些跨时间、跨作品的行为有助于识别更稳定的偏好。For users, the continual focus is on a reliable source of information or a familiar communicator; for recommendation systems, these behaviors across time and across works help in identifying more stable preferences.
4.5 握手模型：留言得到回复的意义4.5 Handshake Model: The Meaning of Receiving a Reply to a Comment
握手模型把用户留言并得到作者回复，视为一次具有积极价值的互动。它关注的是双方形成交流，而不只是两个孤立动作各自发生。The handshake model views a user leaving a comment and receiving a reply from the author as an interaction with positive value. It focuses on the formation of communication between both parties, rather than just two isolated actions occurring separately.
用户提出问题，作者补充细节；用户分享使用经验，作者回应或解释；读者指出疑点，作者修正信息，这些互动都可能让内容的价值继续延伸。作者获得反馈与创作动力，用户感到自己的问题被接住，双方的连接因此增强。A user asks a question, the author provides additional details; a user shares their experience, the author responds or explains; readers point out doubts, the author corrects information—these interactions all have the potential to extend the content's value. Authors receive feedback and creative motivation, while users feel their questions are acknowledged, thus strengthening the connection between both sides.
公开文章还指出，不同作者面对同一条留言的感受可能不同，因此模型需要在更多目标上理解作者与用户双方的体验。这说明作者价值也是推荐生态中的一部分，平台并非只关心观众是否继续观看。Public articles also point out that different authors may feel differently about the same comment, so the model needs to understand both the author's and the user's experiences across more objectives. This indicates that author value is also a part of the recommendation ecosystem; the platform does not only care about whether viewers continue watching.
在实际创作中，回复最有价值的方式通常是推动讨论：回答具体问题、补充遗漏条件、解释做法、感谢有效纠正，或把有代表性的疑问变成后续作品。复制同一句无内容的回复虽然产生了动作，却未必形成同样的交流价值。In actual creation, the most valuable way to reply is usually to drive discussion: answering specific questions, supplementing missing conditions, explaining methods, thanking for effective corrections, or turning representative questions into subsequent works. Simply copying the same content-less reply may create an action, but it does not necessarily generate the same communicative value.
握手模型公开了互动组合的方向，没有公开回复长度、回复速度与流量之间的固定兑换关系。理解这一机制的重点，是作者与观众是否形成有意义的往来。The handshake model reveals the direction of interaction combinations, but does not disclose a fixed conversion relationship between reply length, reply speed, and traffic. The key to understanding this mechanism is whether meaningful exchanges are formed between authors and audiences.
4.6 多目标怎样支持中长视频4.6 How Multi-Objective Supports Medium- and Long-Form Videos
在短视频为主的早期阶段，是否看完是一个直观信号。但当内容长度差异扩大，同一完播率就不再能公平表达消费价值。看完一个很短的视频与认真观看一段长讲解，时间投入和获得的信息量都不同。In the early stage dominated by short videos, whether a video is watched to the end is an intuitive signal. However, when the variation in content length increases, the same completion rate can no longer fairly represent consumption value. Watching a very short video to the end and seriously watching a long explanatory segment involve different time investments and amounts of information gained.
多目标推荐因此同时考虑时长、分享、关注、评论、收藏等反馈。较长的视频即使完播不占优势，也可能在其他目标上表现出较高价值，从而找到更合适的用户。Therefore, multi-objective recommendation simultaneously considers feedback such as duration, shares, follows, comments, and favorites. Even if longer videos do not have an advantage in completion rate, they may show high value in other objectives and thus find a more suitable audience.
公开文章使用“米三汉”450分钟解读《红楼梦》的案例说明这一点：该作品在分享与关注方面表现突出，评论、观看时长和收藏等指标也提供了支持。另一篇案例文章介绍“风雪夜灯”的超长战国历史视频，体现系统化长内容与持续用户需求之间的联系。The public article uses "Mi Sanhan" to interpret "Dream of the Red Chamber" in 450 minutes to illustrate this point: the work excels in sharing and attention, with support from comments, viewing duration, and favorites. Another case article introduces the ultra-long Warring States historical video "Wind and Snow Night Lights," reflecting the connection between systematic long content and ongoing user demand.
长视频的优势是能够容纳复杂解释、完整论证和连续叙事。相应的内容组织也应服务这种优势：开头交代问题与结构，章节之间有清晰递进，例子帮助理解，重复部分适度压缩。仅仅把时长拉长，并不会自动产生深度。The advantage of long videos is their ability to accommodate complex explanations, complete arguments, and continuous storytelling. The corresponding content organization should also serve this advantage: the beginning explains the problem and structure, chapters progress clearly, examples aid understanding, and repetitive parts are moderately compressed. Simply lengthening the length does not automatically create depth.
短内容与长内容的比较，更适合在相似题材、用户需求和消费场景中进行，而不是只比较一个完播率数字。Comparing short and long content is better suited for similar themes, user needs, and consumption scenarios, rather than just comparing a single completion rate figure.
4.7 原创性目标与内容供给的循环4.7 The Cycle Between Originality Goals and Content Supply
原创性也是公开多目标体系中的一个方向。平台鼓励具有独特信息与表达的作品，并对搬运、抄袭等内容限制传播或采取相应处置。其作用不仅是评价当前一条视频，也关系到作者是否愿意继续投入创作。Originality is also one direction within the public multi-objective system. Platforms encourage works with unique information and expression, and restrict or take appropriate measures against content such as copying or plagiarism. Its role is not only to evaluate the current video but also to determine whether the author is willing to continue creating.
目标设计会反过来塑造供给。如果持续表达的作者能够获得关注和追更，更多创作者会重视稳定的人格化表达；如果可保存、可复用的内容能被识别，作者就更有动力制作教程和知识内容。不同价值受到怎样的评价，会影响平台上什么样的作品持续出现。Goal design in turn shapes supply. If authors who continuously express themselves gain attention and follow-ups, more creators will value stable personalized expression; If content that can be saved and reused can be recognized, authors are more motivated to produce tutorials and knowledge content. How different values are evaluated affects which works continue to appear on the platform.
收藏按钮的调整展示了产品与算法的循环：入口更明显，用户更容易准确表达保存意愿；行为信号更清晰，模型更容易学习；有价值的内容因此可能匹配到更多需要它的人，再推动同类优质供给。The adjustment of the favorites button demonstrates the cycle between the product and algorithm: the entry point is clearer, making it easier for users to accurately express their intention to save; Behavioral signals are clearer, making it easier for the model to learn; Valuable content may thus be matched to more people who need it, thereby driving high-quality supply in the same category.
所以，产品设计、反馈数据、模型目标和创作者行为是互相影响的。优质内容生态并不是单靠某个排序公式建立，而是在需求表达与创作回报之间形成长期循环。Therefore, product design, feedback data, model goals, and creator behavior all influence each other. A high-quality content ecosystem is not built solely on a ranking formula, but rather forms a long-term cycle between demand expression and creative returns.
五、推荐系统怎样从行为中学习5. How the Recommendation System Learns from Behavior
5.1 协同过滤：从共同选择中寻找关系5.1 Collaborative Filtering: Finding relationships through shared choices
协同过滤利用许多用户留下的行为，寻找用户之间或内容之间的关联。假设甲看过并喜欢摄影入门、镜头测评和夜景教程，乙喜欢其中前两类，又喜欢胶片冲洗，那么胶片冲洗就可能成为向甲推荐的候选。系统利用的是共同选择所透露的相关性。Collaborative filtering uses the behaviors left by many users to find associations between users or between content. Suppose User A has watched and liked 'Photography for Beginners', 'Lens Reviews', and 'Night Photography Tutorials', and User B likes the first two categories and also enjoys 'Film Development'. Then 'Film Development' could potentially become a candidate recommendation for User A. The system leverages the correlations revealed by common selections.
基于用户的协同过滤先寻找行为相似的人，再参考这些人的选择；基于物品的协同过滤则寻找经常被相同用户消费的内容，再围绕用户已经喜欢的内容扩展。两者都可以在没有人工逐条理解内容的情况下，形成有用的推荐。User-based collaborative filtering first finds people with similar behaviors, then references their choices; item-based collaborative filtering finds content that is frequently consumed by the same users, then expands around content the user already likes. Both methods can generate useful recommendations without manually understanding each piece of content.
为了便于计算，可以把用户作为行、内容作为列，将浏览、评分或其他交互填入矩阵。这张矩阵通常非常稀疏：大多数人只接触全部内容中的极小部分。空白只表示没有相应观测，不能简单等同于讨厌。用户可能从未见过某条内容，而不是看过之后否定了它。For computational convenience, users can be represented as rows and content as columns, with browsing, ratings, or other interactions filled into the matrix. This matrix is usually very sparse: most people have only interacted with a tiny portion of the total content. Blank entries only indicate the lack of corresponding observations and should not be simply equated with dislike. A user may have never seen some content, rather than having seen it and disliked it.
协同过滤能发现人工分类遗漏的联系，但也面对新用户、新内容和小众内容数据不足的问题。一个刚发布的视频还没有共同观看记录，就需要借助内容特征、其他召回方式以及后续反馈获得匹配机会。这也是实际推荐系统通常组合多种方法的原因。Collaborative filtering can discover connections overlooked by manual classification but also faces issues with new users, new content, and niche content due to insufficient data. A newly released video without common viewing records will need features of the content, other retrieval methods, and subsequent feedback to get matching opportunities. This is also why practical recommendation systems usually combine multiple approaches.
5.2 矩阵分解与隐向量5.2 Matrix Factorization and Latent Vectors
矩阵分解尝试把庞大的用户—内容关系压缩成两组较小的表示：一组描述用户，一组描述内容。每个用户和每个内容分别对应一个向量，它们的组合用于估计某种偏好或交互强度。Matrix factorization attempts to compress the vast user-content relationships into two smaller sets of representations: one set describing users and another describing content. Each user and each content item corresponds to a vector, and their combination is used to estimate a type of preference or interaction strength.
一个常见的教学表达是：A common instructional expression is:
其中，(\mathbf p_u) 是用户表示，(\mathbf q_i) 是内容表示，(\hat r_{ui}) 是模型给出的估计值。有些模型还会加入用户偏置、内容偏置和其他项。这个式子说明低维表示怎样参与计算，并不意味着该点积就是最终推荐分数。where \(\mathbf p_u\) represents the user, \(\mathbf q_i\) represents the content, and \(\hat r_{ui}\) is the estimate given by the model. Some models may also include user bias, content bias, and other terms. This formula illustrates how low-dimensional representations participate in the computation and does not imply that this dot product is the final recommendation score.
隐向量中的每个维度不一定对应“美食”“体育”这样的可命名标签。维度通过训练共同承担表达关系的任务，某种偏好可能由多个维度组合表现。因而，用几个固定标签来概括一个用户，通常比模型实际学习到的关系更粗糙。Not every dimension in the hidden vector necessarily corresponds to a nameable label like "food" or "sports." The dimensions collectively take on the task of expressing relationships through training, and a certain preference may be expressed by a combination of multiple dimensions. Therefore, summarizing a user with a few fixed labels is usually coarser than the relationships actually learned by the model.
Embedding是把离散对象或复杂特征映射到向量空间的表示方式。矩阵分解可以学习这种表示，神经网络也可以学习。向量既能服务召回中的相似度搜索，也能成为排序模型的输入；相同的表示思想可以出现在不同环节。Embedding is a way to represent discrete objects or complex features by mapping them into a vector space. Matrix factorization can learn such representations, and neural networks can also learn them. Vectors can serve both for similarity search in retrieval and as inputs to ranking models; the same representation concept can appear in different stages.
5.3 Wide & Deep、特征交叉与行为预估5.3 Wide & Deep, Feature Crosses, and Behavior Prediction
推荐并不只依赖单一特征。喜欢旅行的人未必喜欢所有旅行视频；同一个人下班途中和周末做攻略时，可能需要完全不同的内容。用户偏好、内容特点、时间和场景之间的组合关系，因此十分重要。Recommendations do not rely on a single feature. Someone who likes traveling does not necessarily enjoy all travel videos; the same person commuting after work and planning trips on the weekend may need completely different content. The combination of user preferences, content characteristics, time, and context is therefore very important.
Wide & Deep是一种兼顾记忆与泛化的建模思路。Wide部分有助于利用已经观察到的特征组合，Deep部分通过向量表示与多层网络学习更复杂的关联，让模型能够面对没有完全重复出现过的组合。Wide & Deep is a modeling approach that balances memorization and generalization. The Wide part helps utilize observed feature combinations, while the Deep part, through vector representations and multi-layer networks, learns more complex associations, allowing the model to handle combinations that have not occurred exactly before.
特征交叉就是让模型学习“这些条件一起出现时意味着什么”。例如，喜欢烘焙、近期搜索低糖食谱、周末观看时间较长，这些条件结合起来，比任何一个条件单独出现都更能说明当下需求。组合可以来自人工设计，也可以通过模型结构学习。Feature crossing allows the model to learn "what these conditions mean when they appear together." For example, liking baking, recently searching for low-sugar recipes, and watching longer on weekends—these conditions combined explain current needs better than any single condition alone. Combinations can come from manual design or be learned through the model structure.
训练阶段使用已有行为数据调整参数，让预测结果与观测反馈更一致；推理阶段则把当前用户、候选内容和上下文输入模型，输出各类行为的预测值。输出可能分别对应点赞、收藏、关注等目标，随后再参与价值融合与排序。During training, existing behavior data is used to adjust parameters so that predictions are more consistent with observed feedback; during inference, the current user, candidate content, and context are input into the model to output predictions for various behaviors. The outputs may correspond respectively to likes, favorites, follows, etc., and then participate in value fusion and ranking.
由此需要区分三个层次：特征是输入，参数是训练形成的内部数值，行为概率是模型输出。创作者后台看到的历史互动率与系统针对某个用户预测的下一次行为概率，也不是同一个量。This requires distinguishing three levels: features are inputs, parameters are internally formed values from training, and behavior probabilities are model outputs. The historical interaction rate seen by creators in the backend is not the same quantity as the model's prediction of a user's next action.
六、多样性、长尾兴趣与长期记忆6. Diversity, Long-Tail Interests, and Long-Term Memory
6.1 为什么系统需要主动增加多样性6.1 Why the system needs to actively increase diversity
只持续放大用户最近最强烈的偏好，容易让推荐变得重复。一个人临时看了几条装修视频，并不意味着生活中只剩装修；同一题材大量连续出现，也会降低观看体验。多样性因此同时涉及内容题材、作者、表达方式、消费节奏和兴趣覆盖。Continuously amplifying only the user’s most recent and strongest preferences can easily make recommendations repetitive. Just because someone temporarily watches a few home renovation videos does not mean that their life consists solely of renovations; the repetitive appearance of a single topic can also reduce the viewing experience. Therefore, diversity involves content topics, creators, expression styles, consumption pace, and coverage of interests.
公开页面介绍的做法包括：在推荐结果中打散同类内容，限制相似内容的出现频次，在召回等不同阶段增加多兴趣供给，对长尾内容与新内容进行探索，以及利用搜索和获得授权的社交关系补充信号。The methods introduced on public pages include: mixing similar content in recommendation results, limiting the frequency of similar content appearances, increasing multi-interest offerings at different stages such as recall, exploring long-tail and new content, and supplementing signals through search and authorized social relationships.
探索推荐用于发现“系统还不确定，但用户可能喜欢”的内容。随机探索或对向量表示作适度扰动，可以让候选范围超出当前最强的兴趣区域。探索需要与已有相关性相平衡，否则推荐虽然变化很多，却不能满足用户需求。Exploratory recommendation is used to discover content that "the system is not yet certain about but the user may like." Random exploration or moderate perturbation of vector representations can extend the candidate scope beyond the current strongest interest areas. Exploration needs to be balanced with existing relevance; otherwise, although recommendations may vary greatly, they will not meet user needs.
多样性也不是平均分配所有类别。喜欢历史、园艺和爵士乐的人，不需要每天接收全部题材的等量内容。更合理的方向是：主兴趣得到满足，次兴趣不被长期淹没，潜在兴趣有发现机会，重复内容受到控制。Diversity does not mean evenly distributing all categories. A person who likes history, gardening, and jazz does not need to receive equal amounts of all topics every day. A more reasonable approach is: main interests are fulfilled, secondary interests are not consistently drowned out, potential interests have opportunities to be discovered, and repetitive content is controlled.
从创作角度看，多样性机制为小众内容提供了进入合适人群的路径。它支持的是更丰富的匹配机会，不能据此推出“越冷门越容易获得固定流量奖励”。From the perspective of content creation, diversity mechanisms provide paths for niche content to reach appropriate audiences. They support richer matching opportunities, and one cannot infer from this that "the more niche, the easier it is to receive fixed traffic rewards."
6.2 Trinity：用聚类和直方图记录兴趣6.2 Trinity: Recording Interests Using Clusters and Histograms
Trinity把多元兴趣、长尾兴趣和长期兴趣放在同一个框架中处理。它希望解决三个相互关联的问题：偏好种类记不全，小众爱好抓不准，稳定但低频的兴趣记不久。Trinity handles multi-dimensional interests, long-tail interests, and long-term interests within a single framework. It aims to solve three interrelated problems: incomplete memory of preference types, difficulty capturing niche hobbies, and inability to retain stable but low-frequency interests.
第一个模块是实时兴趣聚类。系统依据内容表示之间的相似性形成内容簇，为这些簇分配标识。簇不必对应人工词典中的单一题材：不同题材的内容可能满足同一种情绪、审美或消费需求，因此在学习到的空间中接近。The first module is real-time interest clustering. The system forms content clusters based on the similarity between content representations and assigns identifiers to these clusters. Clusters do not need to correspond to a single topic in a manual dictionary; content of different topics may satisfy the same emotion, aesthetic, or consumption need, and therefore be close in the learned space.
第二个模块是兴趣直方图。横轴是内容簇，纵轴是用户在较长时间里对相应内容表现出的行为强度。历史观看、充分观看等反馈经过统计，形成相对稳定的兴趣分布。短期偶然多看几条视频，不会立即抹去长期积累的其他偏好。The second module is the interest histogram. The horizontal axis represents content clusters, and the vertical axis represents the intensity of the user’s behavior toward the corresponding content over a long period. Feedback such as historical viewing and fully watched content is statistically aggregated to form a relatively stable interest distribution. Briefly watching several videos in a short period will not immediately erase other preferences accumulated over the long term.
作为教学示意，可以把用户长期观看分布写成摄影50、音乐30、园艺20。近期如果摄影内容占据绝大多数推荐，音乐和园艺就可能是值得检查的欠分发兴趣。这里的数值只是说明统计思路，实际系统还要判断信号是否充分、内容是否适合以及其他推荐条件。As a teaching example, you can write the long-term viewing distribution as photography 50, music 30, gardening 20. If photography content accounts for the vast majority of recommendations recently, music and gardening may be worth checking for underdistributed interest. These values only illustrate the statistical approach; the system also needs to judge whether the signal is sufficient, whether the content is appropriate, and other recommendation criteria.
聚类提供内容之间的结构，直方图提供用户与这些结构之间的历史关系。两者结合，使兴趣不必被压缩成几个固定的文字标签。Clustering provides the structure between content, while histograms provide the historical relationships between users and these structures. By combining the two, interests do not need to be compressed into a few fixed text tags.
6.3 多兴趣、长尾兴趣和长期兴趣分别怎样优化6.3 How to optimize multi-interest, long-tail interest, and long-term interest respectively
多兴趣优化先从长期行为中识别证据充分的兴趣，再检查哪些兴趣获得的分发偏少。对于已经有稳定兴趣表达、但被头部兴趣挤压的内容簇，系统可以增加匹配机会。关键是兴趣可靠且欠分发，而不是只看一个类别最近出现得少。Multi-interest optimization first identifies well-evidenced interests from long-term behavior, then checks which interests receive less distribution. For content clusters that already have stable interest expression but are squeezed out by top interests, the system can increase matching opportunities. The key is that interests are reliable and under-distributed, rather than just looking at a category that has recently appeared less frequently.
长尾兴趣优化需要同时判断内容是否小众、用户是否真正喜欢。网页介绍了一种统计方式：记录内容簇在训练样本中出现的时间间隔，平均间隔较长的类别通常更低频；再把这些低频类别与用户兴趣直方图对照，保留具有显著兴趣证据的部分。Long-tail interest optimization requires judging both whether the content is niche and whether users genuinely like it. The webpage introduces a statistical method: recording the interval between content clusters appearing in training samples; categories with longer average intervals are usually lower frequency; these low-frequency categories are then compared with user interest histograms, retaining parts with significant interest evidence.
这一步把“冷门”与“对某个人有价值”联系起来。偶然划过一次专业摄影视频，与长期反复观看相关内容，所提供的证据强度不同。长期数据有助于区分稳定爱好与偶然接触。This step links "niche" to "valuable to someone." Casually scrolling through a professional photography video and watching related content repeatedly over a long period provide different levels of evidence. Long-term data helps distinguish stable hobbies from casual encounters.
长期兴趣优化则从较久之前的行为中提取代表性内容作为种子，寻找近期相似内容。例如，一个人半年前经常看天文科普，系统可以据此找到新的相关作品，让沉寂一段时间的兴趣重新获得供给。Long-term interest optimization extracts representative content from behaviors from a long time ago as seeds, searching for similar recent content. For example, if someone frequently read astronomy science six months ago, the system can use this to find new related works, resupplying interest that has been dormant for a while.
三类优化互相支持：较长历史揭示多元偏好，多元偏好中包含小众爱好，长期记录又帮助确认小众爱好是否稳定。这也是Trinity强调联合建模的原因。Three types of optimizations support each other: longer history reveals diverse preferences, which include niche hobbies, and long-term records help confirm whether niche interests are stable. This is also why Trinity emphasizes joint modeling.
6.4 STCA怎样读取上万条历史行为6.4 How STCA reads tens of thousands of historical actions
行为序列是按时间排列的交互记录。2025年12月的技术页面将100以下长度作为短序列，将100至10,000量级作为长序列，超过10,000作为超长序列。长既可能表示跨越更久的时间，也可能表示包含更多次行为。Behavior sequences are interaction records arranged in chronological order. The technical page in December 2025 categorizes sequences shorter than 100 steps as short sequences, sequences with magnitudes from 100 to 10,000 as long sequences, and those exceeding 10,000 steps as ultra-long sequences. 'Long' may indicate either a longer time span or more actions included.
传统全量自注意力需要比较序列中各位置之间的关系，主要注意力计算随序列长度呈平方级增长：(O(L^2))。长度增加10倍，对应的两两关系数量约增加100倍。平方增长与指数增长是不同概念。Traditional full self-attention requires comparing relationships between all positions in the sequence, with the main attention calculation growing quadratically with sequence length: (O(L^2)). If the length increases tenfold, the number of pairwise relationships roughly increases by 100 times. Quadratic growth is a different concept from exponential growth.
STCA，即堆叠式目标—历史交叉注意力，把重点转向当前候选内容与历史行为的关系。候选内容形成查询，历史记录提供键和值，通过匹配提取对当前预测有帮助的信息。对于固定数量的目标查询和固定维度，相关注意力计算随历史长度近似线性增长，即(O(L))。STCA, that is, stacked target-historical cross-attention, shifts the focus to the relationship between the current candidate content and historical behaviors. Candidate content forms the query, while historical records provide keys and values, extracting information useful for the current prediction through matching. For a fixed number of target queries and fixed dimensions, the corresponding attention computation grows approximately linearly with historical length: (O(L)).
多层堆叠可以逐步提取复杂关系。与此同时，在同一次推荐请求中，用户历史可以编码一次，再供多个候选内容复用，减少重复计算与数据传输。线性复杂度描述的是这一计算环节，不等于整套服务的所有成本都只取决于历史长度。Multi-layer stacking can gradually extract complex relationships. Meanwhile, in a single recommendation request, the user's history can be encoded once and reused for multiple candidate contents, reducing repeated computation and data transfer. The linear complexity describes this computational step and does not mean that the entire service cost depends solely on historical length.
网页报告，在一万长度的实验场景中，用户历史相关带宽需求降低最高84%，训练吞吐量提升超过2.2倍。页面还介绍了较短序列为主、穿插长序列的训练安排，并在推理时使用更完整历史。其报告约三分之一计算成本达到完整训练80%的效果，但页面没有给出这里“效果”的完整指标定义。The webpage reports that in an experimental scenario with a length of 10,000, the bandwidth demand related to user history was reduced by up to 84%, and training throughput improved by more than 2.2 times. The page also introduces a training arrangement focusing mainly on shorter sequences, interspersed with long sequences, and uses more complete history during inference. About one-third of the computational cost achieves 80% of the effect of full training, though the webpage does not provide the complete definition of "effect" here.
6.5 三项技术在推荐链路中的分工6.5 Division of the Three Technologies in the Recommendation Pipeline
| 技术方向Technical Direction |
主要解决的问题Main Problems Solved |
关键方法Key Methods |
对体验的意义Implications for User Experience |
| Trinity |
多兴趣、小众兴趣和长期兴趣容易被忽略Multi-interest, niche interest, and long-term interest are easily overlooked |
内容聚类、兴趣直方图、联合优化Content clustering, interest histogram, joint optimization |
丰富供给，减少兴趣失忆Enrich supply and reduce interest forgetting |
| STCA |
超长历史难以在可接受成本内使用Ultra-long histories are difficult to use within acceptable cost |
目标—历史交叉注意力、历史复用Goals—Historical Cross-Attention, Historical Reuse |
更充分利用长期行为Make fuller use of long-term behavior |
| RankMixer |
排序模型扩容受到碎片化计算和成本约束Expansion of ranking models is constrained by fragmented computation and costs |
特征Token化、Token Mixing、逐Token网络及稀疏专家Feature tokenization, token mixing, per-token networks, and sparse experts |
提升特征交互与计算效率Improve feature interaction and computational efficiency |
| 大模型推荐探索Exploration of large model recommendations |
传统特征与分阶段处理的表达能力有限Traditional features and staged processing have limited expressive power |
语义理解、序列建模、知识增强及架构探索Semantic understanding, sequence modeling, knowledge enhancement, and architecture exploration |
更好理解复杂内容与需求Better understanding of complex content and requirements |
这些名称描述不同层面的问题，并不构成四套只能择一使用的推荐系统。兴趣组织、历史处理、排序计算和语义理解，可以在复杂系统中形成互补。These names describe issues at different levels and do not constitute four separate recommendation systems from which one must be chosen. Interest organization, historical processing, ranking computation, and semantic understanding can be complementary within a complex system.
还要区分两篇网页中看似相反的策略。RankMixer讨论的是训练与推理时参数模块的激活方式，页面称为“稠密训练—稀疏推理”；STCA讨论的是训练与推理采用的历史序列安排，页面称为“稀疏训练，密集推理”。它们优化的对象不同，不存在直接矛盾。It is also necessary to distinguish between the two web pages that seem to propose opposite strategies. RankMixer discusses the activation of parameter modules during training and inference, referred to on the page as 'dense training—sparse inference'; STCA discusses the arrangement of historical sequences during training and inference, referred to on the page as 'sparse training, dense inference'. They optimize different targets and do not directly contradict each other.
七、大模型与推荐系统的新架构7. Large Models and the New Architecture of Recommendation Systems
7.1 从行为关联走向内容与需求理解7.1 From Behavior Correlation to Content and Demand Understanding
传统推荐能够通过行为关联预测偏好，不必先像人一样读懂每一条内容。大模型则提供了另一类能力：把文本、语义关系和较长上下文纳入分析，帮助系统理解更复杂的内容表达与用户需求。Traditional recommendation can predict preferences through behavior correlation without needing to read each piece of content like a human. Large models, however, provide another type of capability: incorporating text, semantic relationships, and longer context into the analysis to help the system understand more complex content expressions and user needs.
2025年12月的技术文章讨论了两种路径：用大模型增强既有推荐系统，或者探索以大模型为中心重新组织推荐。增强路径可以用于内容理解、用户行为解释和特征表达；更深入的架构探索则尝试减少分阶段处理中的信息损失。公开讨论包含已经开展的实践与下一代方向，不宜把所有研究设想写成全量上线状态。A technical article from December 2025 discussed two paths: using large models to enhance existing recommendation systems, or exploring a recommendation system reorganized with large models as the core. The enhancement path can be used for content understanding, user behavior interpretation, and feature expression; deeper architectural exploration attempts to reduce information loss in staged processing. Public discussions include practices already carried out and next-generation directions, and it is not advisable to present all research concepts as fully deployed.
Transformer的注意力机制让模型按当前任务选择相关信息。其标准教学形式为：Transformers’ attention mechanism allows the model to select relevant information according to the current task. Its standard teaching form is:
Q表示查询，K表示可匹配的键，V表示被聚合的信息，(d_k) 是键向量维度。匹配值经过归一化后决定各部分信息的相对贡献。这个式子解释注意力计算，不是热门分数，也不对应某个固定互动权重。Q represents the query, K represents the matchable key, and V represents the aggregated information. (d_k) is the dimension of the key vector. After normalization, the match values determine the relative contribution of each piece of information. This formula explains attention calculation, not popularity scores, nor does it correspond to a fixed interaction weight.
知识图谱通过实体和关系组织知识，例如作品—作者、景点—城市、食材—菜品等。图谱或其向量表示可以补充内容关联；向量运算表达的是学习到的近似关系，不能把任意坐标相加当作事实推理的充分证明。Knowledge graphs organize knowledge through entities and relationships, such as work—author, attraction—city, ingredient—dish, etc. Graphs or their vector representations can supplement content associations; vector operations express learned approximate relationships and cannot be taken as sufficient proof of reasoning just by arbitrarily adding coordinates.
7.2 RankMixer怎样把零散特征变成高效计算7.2 How RankMixer Turns Scattered Features into Efficient Computation
推荐系统要处理用户、内容、历史行为、发布时间和场景等大量特征。这些输入维度不同、结构不同。如果拆成许多小模块分别计算，即使使用GPU，也可能难以充分发挥并行计算能力。Recommendation systems need to handle a large number of features, including users, content, historical behavior, publication time, and scenarios. These inputs have different dimensions and structures. If calculated separately in many small modules, even with GPUs, the parallel computation capability might not be fully utilized.
RankMixer首先将特征分组，并转换成固定维度的Token表示。这里的Token是模型处理的特征单元，不必对应自然语言中的一个词。统一表示后，计算更容易组织为GPU擅长的大规模矩阵操作。RankMixer first groups features and converts them into fixed-dimension Token representations. Here, a Token is a feature unit processed by the model and does not have to correspond to a word in natural language. Once unified, computation is easier to organize into the large-scale matrix operations that GPUs excel at.
Token Mixing让不同特征单元交换信息，学习组合关系。例如，用户偏好、最近的行为和当前时间共同影响某个视频的适配程度。逐Token前馈网络，即Per-token FFN，则为不同特征单元配置相应的参数处理能力。Token Mixing allows different feature units to exchange information and learn combinational relationships. For example, user preferences, recent behaviors, and the current time all jointly affect a video's suitability. The per-token feedforward network, or Per-token FFN, allocates corresponding parameter processing capacity for different feature units.
网页进一步介绍了稀疏专家混合方案：推理时为当前输入选择部分专家模块，减少不必要的激活与计算。专家是学习到的计算模块，网页中的“手账专家”“旅行专家”属于便于理解的比喻，不应理解为每个主题都配有一个人工命名的固定子模型。The page further explains a sparse expert mixture scheme: during inference, only some expert modules are chosen for the current input to reduce unnecessary activations and computations. Experts are learned computation modules; the "Planner Expert" and "Travel Expert" on the page are metaphors for understanding, and should not be understood as each topic having a fixed submodel with a manually assigned name.
其核心价值是让模型表达能力、参数规模与硬件执行方式更好匹配。更多参数要产生实际收益，还需要有效的数据、训练方法和服务架构支持，参数数量本身并不等于推荐质量。Its core value is that it better aligns model expressiveness, parameter scale, and hardware execution. More parameters will generate actual benefits only with effective data, training methods, and service architecture support; the number of parameters itself does not equal recommendation quality.
7.3 技术页面公布了哪些性能数字7.3 What Performance Numbers Were Published on the Technical Page
以下数值来自2025年12月16日发布的两篇技术网页，适用于页面所述模型与实验背景。The following values come from two technical web pages released on December 16, 2025, applicable to the models and experimental context described on the pages.
| 条目Item |
网页披露的变化Changes Disclosed on the Web Page |
读数含义Meaning of the Reading |
| RankMixer算力利用率MFURankMixer Compute Utilization MFU |
4.5%提升到45%以上Increased from 4.5% to over 45% |
计算资源的有效利用程度提升Improvement in the effective utilization of computing resources |
| RankMixer推理核心活跃度RankMixer Inference Core Activity |
30%提升到80%Increased from 30% to 80% |
页面所述核心活跃指标提升Improvement in core activity metrics described on the page |
| 排序模型参数规模Ranking Model Parameter Scale |
1,600万提升到11亿，约70倍Increased from 16 million to 1.1 billion, about 70 times |
模型容量显著扩大Significant expansion of model capacity |
| 上述参数扩容的推理成本Inference cost for the above parameter expansion |
页面称未增加Stated on the page as unchanged |
对应所述架构与比较条件Corresponding architecture and comparison conditions |
| STCA一万长度行为序列STCA 10,000-length behavioral sequence |
历史相关带宽需求最高降低84%Maximum reduction in historical relevance bandwidth demand by 84% |
减少历史处理中的数据传输需求Reduces data transfer requirements in historical processing |
| STCA训练吞吐量STCA Training Throughput |
提升超过2.2倍Increased by more than 2.2 times |
单位时间可完成的训练处理增加Increased training processing completed per unit of time |
这些指标回答的是模型如何更高效地处理数据。它们与单个视频的播放量、创作者收入、用户满意度属于不同指标体系。例如，模型参数扩大约70倍，不表示作品流量扩大约70倍；带宽下降84%，也不意味着所有服务器费用同步下降84%。These metrics answer how the model handles data more efficiently. They belong to a different metric system from single video views, creator revenue, or user satisfaction. For example, expanding model parameters by about 70 times does not mean content traffic expands by about 70 times; reducing bandwidth by 84% does not mean all server costs decrease by 84% simultaneously.
从应用意义看，更高的计算效率为更长历史、更复杂特征和更大的模型腾出空间。最终体验仍要由匹配质量、多样性、长期使用反馈和内容环境共同评价。In practical terms, higher computational efficiency allows for longer histories, more complex features, and larger models. The final experience still needs to be evaluated by a combination of relevance quality, diversity, long-term usage feedback, and content environment.
八、内容从发布到高热的审核与质量判断8. Review and Quality Assessment from Publishing Content to Going Viral
8.1 审核之前，先定义什么问题需要处理8.1 Define what issues need to be addressed before review
内容治理首先需要把原则转化为可执行标准。案例平台设置标准管理团队，负责标准的纳入、退出、适用范围与更新。2025年1月的页面介绍了十余类、数百种风险条款，并为条款配套典型案例和执行手册。Content governance first requires converting principles into executable standards. The case platform sets up a standards management team responsible for the inclusion, removal, scope, and updating of standards. The page in January 2025 introduced more than ten categories and hundreds of risk clauses, along with typical cases and implementation manuals for each clause.
标准既按问题分类，也按载体和场景区分。视频、图片、文字的判断方式不同，昵称、评论、话题等产品位置的使用情境也不同。相同的一段文字放在不同场景中，其指向和影响可能发生变化。Standards are categorized both by problem type and by medium and scenario. The judgment methods for videos, images, and text differ, and the usage context for product positions such as nicknames, comments, and topics also varies. The same piece of text may have different intent and impact depending on the scenario in which it appears.
标准更新包括常规评估和突发问题的快速响应。变更通过线上流程与审批留痕，便于追踪版本；质量管理组检查执行情况，培训组帮助执行团队统一理解并纠正偏差。对用户和作者的规则说明、违规通知，则把执行结果连接到具体使用体验。Standard updates include routine evaluations and rapid responses to emergent issues. Changes are recorded and approved through online processes for traceability; the quality management team checks implementation, and the training team helps the execution team unify understanding and correct deviations. Rules explanations and violation notifications for users and creators link execution results to specific user experiences.
这套体系提供了机器识别与人工判断共同遵循的基础。模型并不是自行决定全部规则，而是在既定标准、样本和持续反馈中学习如何识别问题。内容审核因此同时包含规则工作、技术工作和执行质量管理。This system provides a foundation for both machine recognition and human judgment to follow. Models do not autonomously decide all rules; they learn how to identify problems within established standards, samples, and continuous feedback. Therefore, content review includes rules work, technical work, and execution quality management.
8.2 初审、复审、三审和高热质量研判8.2 Initial Review, Re-review, Third Review, and High-Engagement Quality Assessment
初审发生在上传后的识别阶段，重点检查底线风险。命中高危特征的内容可能直接被拦截；模型判断较大概率存在底线问题的内容进入人工研判；风险判断较低的内容可以获得基础流量，继续进入分发环节。The initial review occurs at the recognition stage post-upload, focusing on bottom-line risks. Content hitting high-risk features may be directly blocked; content that the model judges as likely having bottom-line issues enters human review; content with lower risk assessment can receive basic traffic and continue to the distribution stage.
复审发生在内容达到一定播放量级后，检查是否存在不适合过多传播的问题，例如画面容易引起不适。未命中相关标准的内容继续自然推荐，命中问题标准的内容停止进一步分发。Re-review happens after the content reaches a certain play volume, checking whether there are issues unsuitable for widespread dissemination, such as images that may cause discomfort. Content not hitting relevant standards continues natural recommendation, while content that hits problem standards stops further distribution.
三审面向播放量继续增长的内容，判断逻辑类似，但标准更严格。达到更高热度后，运营人员还会综合研判内容质量，判断是否适合继续传播。整个过程体现了从底线判断，到扩大传播适宜性，再到高热内容质量的逐层考察。Third review targets content whose play volume continues to grow, following similar judgment logic but with stricter standards. Once higher engagement is reached, operations personnel also comprehensively assess content quality to determine if it is suitable for continued dissemination. The entire process reflects a tiered examination from bottom-line judgment, to expansion appropriateness for distribution, and then to high-engagement content quality.
| 环节Stage |
主要关注Main focus |
与分发的关系Relationship with distribution |
| 初审Preliminary Review |
底线风险、高危特征Baseline Risk, High-Risk Features |
影响内容能否开始进入推荐Affects whether content can start entering recommendations |
| 复审Secondary Review |
扩大传播可能引发的不适等问题Discomfort or other issues that may be triggered by wider dissemination |
影响是否继续自然分发Affects whether distribution continues naturally |
| 三审Tertiary Review |
更高传播量下的严格判断Strict judgment for higher levels of dissemination |
影响后续传播Affects subsequent dissemination |
| 高热质量研判High-Engagement Quality Assessment |
原创、优质及传播适宜性Originality, quality, and suitability for dissemination |
影响高热内容是否继续扩大Affects whether high-engagement content continues to expand |
原文明确说明播放量级阈值处于波动状态，没有公布统一的500、3,000、1万等固定台阶。审核节点也不等于推荐算法的全部分发层级，不能把这篇文章改写成一张固定流量池晋级表。The original text clearly states that the playback volume threshold is fluctuating and that a fixed step such as 500, 3,000, or 10,000 has not been announced. Review nodes do not equal all distribution layers of the recommendation algorithm, so this article cannot be rewritten into a fixed traffic pool advancement table.
8.3 审核还会被哪些信号触发8.3 What other signals can trigger reviews
播放量增长只是触发治理判断的一类信号。用户举报、评论区集中质疑、流量异常激增，也可能触发新的机器识别与人工研判。一个视频可能在多个位置进入审核，实际路径并不总是按三个步骤单向前进。Playback growth is only one type of signal that triggers governance judgments. User reports, concentrated questioning in the comment section, and unusual traffic surges may also trigger new machine recognition and manual review. A video may enter review at multiple points, and the actual path does not always move linearly through the three steps.
机器主要承担广泛扫描和快速筛选，人工则处理疑难语境、复杂事实和高热内容的综合判断。人工还通过样本、标准调整与误判分析持续校准模型。两者的关系是覆盖面与判断深度互相补充。Machines mainly handle broad scanning and rapid filtering, while humans handle difficult contexts, complex facts, and comprehensive judgment of high-engagement content. Humans also continuously calibrate the model through samples, standard adjustments, and analysis of misjudgments. The relationship between the two is one of coverage and judgment depth complementing each other.
治理体系的迭代也不止是删除一条问题内容。新问题出现后，需要检查标准是否覆盖、识别是否有效、处置策略是否适合、用户反馈是否改善。相关能力升级时，还会检查存量内容，减少旧问题继续存在的情况。The iteration of the governance system is not just deleting a problematic piece of content. When new issues arise, it is necessary to check whether standards are comprehensive, identification is effective, disposal strategies are appropriate, and user feedback is improved. When capabilities are upgraded, existing content is also reviewed to reduce the continued presence of old problems.
因此，播放量变化可能来自多种环节：用户匹配变化、候选竞争、重复疲劳、质量评估或具体治理处置。单看播放曲线无法确定是哪一种原因。对创作者更有用的证据是作品状态、平台通知、内容实际问题和受众反馈之间能否相互印证。Therefore, changes in playback volume may come from multiple links: changes in user matching, candidate competition, repetition fatigue, quality evaluation, or specific governance measures. Observing the playback curve alone cannot determine which reason it is. More useful evidence for creators is whether the status of the work, platform notifications, actual content issues, and audience feedback can corroborate each other.
九、评论质量怎样影响推荐9. How Comment Quality Affects Recommendations
9.1 评论数量之外，系统还判断评论价值9.1 Beyond the Number of Comments, the System Also Assesses Comment Value
2026年5月的评论治理页面介绍了正向评论目标和恶评目标。系统不仅预测一条内容是否会引发评论，也尝试预测它会引发怎样的评论，并将这些判断纳入多目标体系。The comment governance page from May 2026 introduced the goals for positive comments and the targets for negative comments. The system not only predicts whether a piece of content will attract comments but also tries to predict what kind of comments it will generate, incorporating these judgments into a multi-objective framework.
正向评论可以是鼓励、经验分享、观点表达、客观信息、求助提问、合适的幽默或兴趣交流。共同特征是为参与者创造交流价值。正向并不意味着每个人都同意作者，也不要求评论只能表扬。Positive comments can include encouragement, sharing experiences, expressing opinions, providing objective information, asking for help, appropriate humor, or exchanging interests. The common feature is creating value for participants in the conversation. Positive does not mean everyone agrees with the author, nor does it require comments to be purely praising.
恶评目标关注人身攻击、引战对立、低俗冒犯、垃圾广告和恶意刷屏等反馈。对恶评概率与内容氛围的判断，可以减少相应曝光，限制依赖攻击性互动获得传播的内容。Negative comment targets focus on personal attacks, conflict incitement, vulgar offense, spam advertising, and malicious flooding. Judgments about the likelihood of negative comments and the content atmosphere can reduce corresponding exposure and limit content that relies on aggressive interactions to gain visibility.
| 同样出现评论Comments also appear |
更有价值的情形More valuable situations |
需要区分的情形Situations that need to be distinguished |
| 提出不同意见Expressing a different opinion |
给出理由、事实和具体建议Providing reasons, facts, and specific suggestions |
对人进行羞辱或煽动围攻Humiliating people or inciting mob attacks |
| 分享经历Sharing experiences |
补充真实体验与使用条件Adding real experiences and usage conditions |
无关广告或重复灌水Irrelevant advertisements or repeated spam |
| 幽默与玩梗Humor and memes |
与语境相符，帮助交流Contextually appropriate, helping communication |
借梗隐晦攻击当事人Using memes to subtly attack involved parties |
| 专业讨论Professional discussion |
指出细节、方法与证据Pointing out details, methods, and evidence |
用身份压人、无依据贬损Oppressing people with identity or disparaging without basis |
这意味着评论多和评论好需要分开观察。引战可能制造很高的短期互动量，但这类互动进入质量与恶评目标后，不会天然转化为持续分发优势。This means that the quantity and quality of comments need to be observed separately. Controversy may generate high short-term interaction, but such interaction, once evaluated under quality and negative comment objectives, will not naturally convert into sustained distribution advantage.
9.2 大模型怎样理解评论的言外之意9.2 How Large Models Understand the Implicit Meaning of Comments
同一句话在不同视频下可能表达相反态度；文字与图片分别普通，组合起来也可能具有攻击性。单纯匹配关键词，难以识别反话、指代、表情组合和不断变化的网络表达。The same sentence under different videos may express opposite attitudes; text and images individually may be ordinary, but when combined, they can also be aggressive. Simply matching keywords makes it difficult to identify sarcasm, references, emoji combinations, and continuously evolving online expressions.
案例平台的评论治理页面介绍，2025年开始接入大模型能力，利用更强的上下文与语义分析判断评论含义。模型需要结合原视频、评论内容及相关语境，区分合理建议、善意调侃与恶意攻击。The case platform's comment governance page introduces that starting in 2025, large model capabilities will be integrated to use stronger contextual and semantic analysis to determine the meaning of comments. The model needs to consider the original video, the comment content, and the relevant context to distinguish between reasonable suggestions, good-natured teasing, and malicious attacks.
训练方法包括人类反馈强化学习RLHF和直接偏好优化DPO。前者利用人工评价形成奖励信号，推动模型倾向于更好的解读；后者直接利用成对的较优和较差解读来优化参数。例如，“建议讲慢一点”可以是具体改进意见，不应因为包含不满就自动判为攻击。Training methods include Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO). The former uses human evaluations to form reward signals, promoting the model to interpret better; the latter directly uses paired better and worse interpretations to optimize parameters. For example, "suggest speaking a little slower" can be a specific improvement suggestion and should not automatically be considered an attack just because it contains dissatisfaction.
在线处理则采用大小模型协同：较明确的常见问题交给轻量模型，依赖深层语境的表达交给能力更强的模型。页面还介绍了让Agent分析案例、由专家评审并纠偏的探索方向。Online processing adopts a collaboration between large and small models: clear and common issues are handled by lightweight models, while expressions requiring deep context rely on more capable models. The page also introduces exploratory directions such as having agents analyze cases, reviewed and corrected by experts.
这种改进既服务评论本身的处置，也服务内容氛围与推荐目标。创作者可以从中得到一个直接启示：围绕问题、经验和证据设计讨论，比只追求一句话激起更多情绪更符合长期内容价值。These improvements serve both the handling of comments themselves and the content atmosphere and recommendation goals. Creators can gain a direct insight from this: designing discussions around problems, experience, and evidence aligns better with long-term content value than simply trying to provoke more emotions with a single sentence.
十、谣言识别、AI内容与网络暴力治理10. Rumor Detection, AI Content, and Online Violence Governance
10.1 AI求真先判断哪些内容值得核查10.1 AI Truth-Seeking First Determines Which Content Is Worth Verifying
AI求真面向具有公共价值、现实误导风险且能够核实的资讯。个人观点、明显玩笑、夸张修辞和剧情表达，与需要查证的事实陈述属于不同任务。把所有内容一律送去真假裁决，既浪费计算，也容易误解正常表达。AI truth-seeking targets information with public value, real-world misleading risks, and verifiability. Personal opinions, obvious jokes, exaggerated rhetoric, and narrative expressions belong to different tasks than factual claims that need to be verified. Sending all content for a true-or-false judgment is both a waste of computing resources and likely to misinterpret normal expressions.
公开的处理链路因此设置了前置筛选。谣言领域大模型依据题材、体裁和历史经验，寻找更可能存在问题的内容；风险过滤模块进一步判断公共性风险，并过滤容易被误伤的低风险内容。The public handling workflow therefore sets a pre-screening. In the rumor domain, the large model identifies content more likely to have issues based on topic, genre, and historical experience; the risk filtering module further assesses public risk and filters out low-risk content that could easily be mistakenly flagged.
风险过滤结合了从历史误判案例训练的小模型，以及判断公共性风险的大语言模型。两个方向配合，让后续研判更集中于需要查证的候选内容。Risk filtering combines a small model trained on historical misjudgment cases with a large language model that assesses public risk. The cooperation of the two approaches allows subsequent evaluation to focus more on candidate content that needs verification.
经过筛选后，求真系统再提取具体主张，例如某件事是否发生、时间地点是否正确、视频中的对象是否与标题一致。主张越明确，越容易组织检索和证据比对。After screening, the fact-checking system extracts specific claims, such as whether a certain event occurred, whether the time and place are correct, and whether the objects in the video match the title. The clearer the claim, the easier it is to organize retrieval and evidence comparison.
这套分工也说明，召回候选与最终判断是两件事。被筛出来接受核查，只代表存在进一步检查的必要，不等于已经被认定为谣言。This division of labor also shows that recalling candidates and making the final judgment are two separate matters. Being screened and accepted for verification only indicates the necessity for further examination, and does not mean it has already been identified as a rumor.
10.2 AI求真怎样利用检索与资料库10.2 How AI Fact-Checking Uses Retrieval and Databases
AI求真的关键能力之一，是在判断前寻找外部证据。系统结合内部谣言资料库与网络搜索，查询已核实案例、官方通告和相关新闻，并根据问题进展展开多轮检索。One of the key capabilities of AI fact-checking is to search for external evidence before making a judgment. The system combines an internal rumor database with web search to query verified cases, official announcements, and related news, and conducts multiple rounds of retrieval based on the progress of the inquiry.
这一过程属于检索增强生成RAG的应用：模型先获得与问题相关的材料，再基于材料组织判断。它不同于只凭模型训练记忆直接回答。资料是否新、是否相关、是否支持当前主张，都会影响判断质量。This process belongs to the application of retrieval-augmented generation (RAG): the model first obtains materials related to the question, then organizes judgments based on the materials. It differs from answering solely based on the model's learned memory. Whether the data is new, relevant, or supportive of the current claim will affect the quality of the judgment.
网页介绍了信源分级机制。政务发布平台、中央新闻单位等在其体系中具有较高优先级，论坛帖文、自媒体等信息需要结合其他证据核验。不同来源出现矛盾时，需要比较发布时间、事件对象和原始依据，而不是只数有多少网页重复同一句话。Webpages introduce a source grading mechanism. Government publishing platforms, central news agencies, and similar sources have higher priority in this system, while forum posts and personal media require verification with other evidence. When contradictions appear among sources, it is necessary to compare publishing times, event subjects, and original references, rather than simply counting how many webpages repeat the same sentence.
内部资料库还能帮助发现旧谣言变体：更换地点、日期、产品名称或叙述方式，可能保留相同的核心误导结构。相似性可以触发疑似标记与进一步核查，但一个既有案例并不能自动证明所有相近的新主张都为假。The internal database can also help identify old rumor variants: by changing the location, date, product name, or narrative style, the same core misleading structure may be retained. Similarity can trigger suspicion marks and further verification, but an existing case does not automatically prove that all similar new claims are false.
对于知识库建设，这一案例展示了元数据的重要性：记录来源、日期、事件对象和适用范围，可以帮助检索结果被正确解释。只保存脱离上下文的一段结论，容易把旧事实套用到新问题。For knowledge base construction, this case demonstrates the importance of metadata: recording the source, date, event subject, and applicable scope can help retrieval results be interpreted correctly. Only saving a conclusion out of context can easily lead to the application of old facts to new questions.
10.3 核查结果怎样改变传播10.3 How Verification Results Affect Dissemination
事实清楚、证据明确的问题内容可以进入自动化处置；疑似内容可以降低推荐权重、挂载风险提示；难以确定的内容需要进一步处理。涉及公共安全、自然灾害、金融市场等高风险领域时，页面介绍了人工协同复核安排。Content with clear facts and definite evidence can enter automated handling; suspected content can have its recommendation weight reduced and be flagged with risk alerts; content that is difficult to determine requires further processing. When involving high-risk areas such as public safety, natural disasters, and financial markets, the page introduces arrangements for collaborative human review.
谣言治理还包括澄清信息的传播。平台可以向看过不实内容的人匹配相关辟谣材料，帮助纠正已经形成的印象，而不只停止原内容继续传播。Rumor management also includes the dissemination of clarifying information. Platforms can match relevant debunking materials to people who have seen false content, helping to correct already formed impressions instead of just stopping the original content from continuing to spread.
2026年4月的案例网页以广西救护车争议为例，介绍平台在核实后结合相关搜索和定向推荐传播澄清信息。这个案例的重点是：事实核查、搜索呈现、内容推荐和后续纠偏共同构成传播治理。The April 2026 case webpage uses the Guangxi ambulance controversy as an example, introducing how the platform, after verification, combines related searches and targeted recommendations to spread clarifying information. The focus of this case is: fact-checking, search presentation, content recommendation, and subsequent correction together constitute dissemination governance.
2025年度报告网页公布，谣言曝光量下降90%，用户举报量下降74%，辟谣提醒日均覆盖约3亿次。这些是报告所述阶段的治理数据；曝光、举报和提醒次数衡量的是不同环节，不宜合并成同一个“准确率”。The 2025 annual report webpage published that rumor exposure decreased by 90%, user reports decreased by 74%, and debunking reminders reached about 300 million times per day on average. These are governance data for the reported period; exposure, reports, and reminders measure different stages and should not be combined into a single 'accuracy rate.'
10.4 AI内容识别使用哪些线索10.4 What cues AI content recognition uses
案例平台对AI创作的基本方向是支持创意表达，同时提高生成内容的可识别性。AI标识帮助用户理解内容的制作方式，识别误导性则关注它是否被包装成真实事件、真实身份或真实记录。The case platform’s basic approach to AI creation is to support creative expression while increasing the recognizability of generated content. AI labels help users understand how content is produced, while misleadingness identification focuses on whether it is packaged as a real event, real identity, or real record.
识别线索包括创作者主动标注、显式水印、隐式水印，以及模型对文本、图片、音频和视频的分析。不同来源的信号可以互相补充，不能只依赖某一种生成工具留下的固定特征。Recognition cues include creators’ voluntary labels, explicit watermarks, implicit watermarks, and model analysis of text, images, audio, and video. Signals from different sources can complement each other and cannot rely solely on fixed features left by a single generation tool.
多模态模型结合视频画面、标题等信息作综合判断；对口型检测关注人物口型与声音的对应关系；音频模型寻找合成声音特征；误导倾向模型则判断内容是否容易让用户产生事实混淆。Multimodal models make comprehensive judgments by combining video frames, titles, and other information; lip-sync detection focuses on the correspondence between a person’s lip movements and sound; audio models look for synthesized voice characteristics; misleadingness models determine whether content is likely to cause factual confusion for users.
“是否AI生成”和“是否误导公众”是两个不同问题。一部明确标注的幻想短片可以具有创作价值，一条以真实新闻口吻发布的合成灾害画面则涉及事实误导。两类判断需要在内容语境中结合使用。"Whether AI-generated" and "whether misleading to the public" are two different issues. A clearly labeled fantasy short film can have creative value, while a synthetic disaster image presented in the tone of real news involves factual misrepresentation. Both types of judgments need to be combined within the content context.
2026年6月的页面把SLNet作为高逼真图像和视频检测的一项技术，进一步介绍其频域与局部伪影分析方法。The June 2026 webpage introduces SLNet as a technology for detecting highly realistic images and videos, further explaining its frequency domain and local artifact analysis methods.
10.5 SLNet怎样寻找生成痕迹10.5 How SLNet searches for generation traces
图像的空间域表示描述像素在画面上的分布，频域表示则描述明暗、纹理等变化的频率结构。快速傅里叶变换FFT可以把图像转换到频域，帮助分析肉眼不容易直接看出的统计规律。The spatial domain representation of an image describes the distribution of pixels on the frame, while the frequency domain representation describes the frequency structure of variations such as brightness and texture. The Fast Fourier Transform (FFT) can convert an image to the frequency domain, helping analyze statistical patterns that are not easily seen with the naked eye.
平滑区域主要对应较低频率，边缘、纹理和细小变化包含较高频率。生成过程、上采样等环节可能形成特定频谱模式；真实拍摄也会受到镜头、传感器、压缩和后期处理影响。因此，检测依赖学习与组合判断，不是看到某个高频峰值就能直接断言为AI。Smooth areas mainly correspond to lower frequencies, while edges, textures, and fine details contain higher frequencies. Generation processes, upsampling, and other steps may create specific frequency patterns; real photography is also affected by lenses, sensors, compression, and post-processing. Therefore, detection relies on learning and combined judgment, and one cannot directly assert an image is AI-generated just by seeing a certain high-frequency peak.
SLNet的另一条路径是局部伪影，例如异常重复的纹理、不自然的边缘过渡、局部结构或光照关系不协调。随着生成质量提高，明显的手指错误等线索变少，检测更需要关注细微的统计异常。Another approach of SLNet is local artifacts, such as abnormal repeated textures, unnatural edge transitions, and inconsistent local structures or lighting relationships. As generation quality improves, obvious cues like finger errors decrease, making detection more reliant on subtle statistical anomalies.
为降低对题材和画风的依赖，团队构造语义接近的真实与生成样本，使模型无法只靠“画的是什么”区分真假。内容尽量相同、制作痕迹不同的困难样本，有助于模型学习更有泛化能力的线索。To reduce dependence on subject matter and style, the team constructs real and generated samples with similar semantics so that the model cannot distinguish authenticity just based on “what is being depicted.” Difficult samples where content is as similar as possible but traces of creation differ help the model learn more generalizable cues.
频域、伪影、语义与其他检测信号构成互补。它们反映检测系统的技术方向，并不构成一个人人都能用肉眼执行、对所有新模型都有效的真假鉴定公式。Frequency domain, artifacts, semantics, and other detection signals form a complementary system. They reflect the technical direction of the detection system and do not constitute a formula for authenticity that anyone can execute with the naked eye or that works for all new models.
10.6 AI反网暴如何从零散互动发现风险10.6 How AI Counter Cyberbullying Identifies Risks from Scattered Interactions
网络暴力可能以连续、低频、隐晦的互动累积形成。只检查一句话是否包含辱骂，容易遗漏围绕同一个人的持续攻击。2026年6月的页面介绍了“识别模型＋研判模型”的双层体系。Online violence may accumulate through continuous, low-frequency, and subtle interactions. Simply checking whether a sentence contains abuse is likely to miss sustained attacks on the same person. A page introduced in June 2026 describes a two-layer system of “recognition model + judgment model.”
第一层利用轻量模型广泛寻找线索，包括评论、提及、陌生人私信突然增加，当事人自述遭受网暴，大量删评或频繁举报等变化。模型给出风险评分，把潜在对象送入后续研判。Agent分析案例与误判，帮助更新识别特征和模型。The first layer uses lightweight models to widely search for clues, including comments, mentions, sudden increases in private messages from strangers, self-reports of suffering from cyberbullying, mass comment deletions, or frequent reports. The model outputs risk scores and sends potential cases to subsequent judgment. Agents analyze cases and misjudgments to help update recognition features and models.
第二层使用多模态大模型，结合文本、图片、音频、视频与统计特征恢复语境。OCR提取图片文字，语音识别提取音频信息，视频总结帮助梳理事件。较长时间范围内的互动和行为变化，使零散风险能够被联系起来。The second layer uses multi-modal large models, combining text, images, audio, video, and statistical features to restore context. OCR extracts text from images, speech recognition extracts audio information, and video summaries help organize events. By looking at interactions and behavior changes over a longer period, scattered risks can be connected.
研判不仅输出是否存在风险，还可以形成事件小结、风险来源和关键证据，供人工复核及后续保护使用。规则、典型案例、人工标注与模型训练共同支撑这一流程。The analysis not only assesses whether there is a risk in the output, but can also generate event summaries, sources of risk, and key evidence for manual review and subsequent protection use. Rules, typical cases, manual labeling, and model training jointly support this process.
页面还强调对不同人群与场景作差异化判断，并衔接防护工具、人工联系和援助。技术的作用从识别单条攻击内容，扩展到发现需要帮助的用户。The page also emphasizes making differentiated judgments for different groups and scenarios and connecting protective tools, manual contact, and assistance. The role of technology extends from recognizing individual offensive content to discovering users who need help.
10.7 网暴治理数据应怎样读取10.7 How to Read Cyberbullying Governance Data
2026年6月12日发布的AI反网暴文章披露，升级后日均筛选的潜在受害者范围扩大91.5%，获得人工外呼帮助的用户数量增长147%。截至文章发布时，其所述“2026年以来”预警识别5,311例涉嫌网暴风险事件，拦截4,294万条攻击类评论，向324人提供免费法律援助。An AI anti-cyberbullying article published on June 12, 2026, disclosed that after the upgrade, the daily screening range of potential victims expanded by 91.5%, and the number of users receiving manual outreach help increased by 147%. As of the article's publication, it stated that since 2026, there were 5,311 suspected cyberbullying risk incidents identified, 42.94 million offensive comments blocked, and 324 people provided with free legal aid.
2025年度报告网页则公布了另一统计阶段的数据：全年拦截不当评论约1.2亿条，向约132万个账号定向推送一键防网暴提醒；2025年1月至11月为635名用户提供免费法律咨询。The 2025 annual report webpage published data from another statistical period: about 120 million inappropriate comments were blocked throughout the year, automated anti-cyberbullying reminders were sent to approximately 1.32 million accounts, and free legal consultations were provided to 635 users from January to November 2025.
这些数字分别描述筛选范围、服务触达、风险事件、评论处置和援助服务。风险事件不等于最终确认的违法案件，评论条数不等于人数，法律咨询与法律援助也不是同一服务口径，不宜直接相加比较。These figures describe screening coverage, service reach, risk incidents, comment handling, and aid services, respectively. Risk incidents are not equivalent to finally confirmed illegal cases, the number of comments is not equal to the number of people, and legal consultation and legal aid are not the same service standard, so they should not be directly added or compared.
从平台能力角度，这些数据呈现的重点是前置发现与后续帮助的连接。知识库保存时应保留发表日期、统计期间和计量单位，使后续检索能回答“哪一年、截至什么时候、统计的是什么”。From the platform capability perspective, the focus of these data is the link between early detection and subsequent assistance. When storing in a knowledge base, the publication date, statistical period, and measurement units should be preserved, so that future searches can answer 'which year, up to when, and what is being measured.'
十一、用户怎样调整自己看到的内容11. How Users Adjust the Content They See
11.1 使用管理助手与内容偏好管理11.1 Using the Management Assistant and Content Preference Management
使用管理助手把观看时长分析与内容偏好管理放在一起。2025年5月的介绍给出的入口是首页侧边栏或“我”的主页侧边栏，再进入更多功能。2026年面向长辈的网页指南继续介绍了这一工具。The management assistant combines viewing time analysis with content preference management. The introduction in May 2025 provided access through the homepage sidebar or the “Me” homepage sidebar, then into more functions. The 2026 web guide for elderly users continued to introduce this tool.
内容偏好圆环图按照电影、美食、旅行等类别，呈现近期观看占比。它帮助用户观察自己最近在看什么；图中的分类用于产品展示，不代表底层推荐只依靠这些固定文字标签。The content preference donut chart displays the proportion of recent viewing by categories such as movies, food, and travel. It helps users observe what they have recently been watching; the categories in the chart are used for product display and do not mean that the underlying recommendations rely solely on these fixed text labels.
“探索更多”会适当减少头部兴趣的重复，增加用户较少接触但可能感兴趣的内容。它的目标是扩大兴趣覆盖，同时维持相关性，而不是把全部历史偏好清空。"'Explore More' will appropriately reduce repetition of head interests and increase content that users have less exposure to but may be interested in. Its goal is to expand interest coverage while maintaining relevance, rather than clearing all historical preferences.
“调节内容偏好”提供减少、默认、增加三个档位，用户可以对类别表达明确偏好。屏蔽设置则用于添加不想看到的内容。选择之后，系统会结合这些反馈调整推荐，后续实际浏览行为仍会继续参与学习。'Adjust Content Preferences' offers three levels: reduce, default, and increase, allowing users to express clear preferences for categories. The block settings are for adding content you don’t want to see. After selection, the system adjusts recommendations based on this feedback, and subsequent actual browsing behavior will continue to contribute to learning.
| 用户想解决的问题Problems users want to solve |
对应操作Corresponding actions |
| 想知道最近看了哪些类型Want to know what types were recently watched |
查看近期观看占比Check the proportion of recent views |
| 推荐过于单一Recommendations are too monotonous |
使用探索更多，并主动接触新主题Use 'Explore More' and actively engage with new topics |
| 某个类别过多或过少Too much or too little of a certain category |
调整类别偏好档位Adjust category preference levels |
| 某类内容明确不想看Clearly do not want to watch a particular type of content |
使用屏蔽设置或不感兴趣Use the block settings or 'Not Interested' |
| 想管理使用时间Want to manage usage time |
查看时长分析并使用相应提醒功能Check duration analysis and use the corresponding reminder features |
11.2 不感兴趣、搜索和日常行为如何配合11.2 How 'Not Interested,' Search, and Daily Behavior Work Together
刷到不想继续看的内容，可以及时划过；希望更明确表达排斥时，可以长按视频或从分享菜单选择不感兴趣。页面说明，这类反馈会减少相似内容的推荐。面向长辈的指南还介绍了屏蔽特定作者以及设置不想看的内容。If you come across content you don’t want to continue watching, you can swipe past it immediately; if you want to express rejection more clearly, you can long-press the video or select 'Not Interested' from the share menu. The page explanation states that this type of feedback will reduce recommendations of similar content. The guide for older users also introduces how to block specific creators and set content they do not want to see.
想增加某种内容时，可以主动搜索、观看真正需要的作品，并进行真实的点赞、收藏、关注和分享。搜索是较明确的主动需求表达，搜推联动能够把这种需求补充到后续推荐中。When wanting to increase a certain type of content, you can actively search for and watch works you really need, and give genuine likes, bookmarks, follows, and shares. Searching is a more explicit expression of active demand, and search-recommendation interaction can supplement this demand into future recommendations.
几类动作表达的含义有所不同：观看说明投入注意，点赞表达即时认可，收藏表示希望保存，关注表达持续接收作者内容的意愿。组合起来，系统更容易区分短暂好奇与长期需要。The meanings of several actions differ: watching indicates attention investment, liking expresses immediate approval, bookmarking means wanting to save, and following expresses willingness to continue receiving content from the creator. Together, the system can more easily distinguish between short-term curiosity and long-term need.
如果用户嘴上不喜欢某类内容，却持续长时间观看并参与争论，行为中就会出现相互冲突的信号。明确设置偏好，同时改变实际消费方式，有助于更清楚地表达需求。If users say they dislike a certain type of content but continue watching and engaging in debates for a long time, conflicting signals will appear in their behavior. Clearly setting preferences while changing actual consumption patterns helps express needs more clearly.
对创作者而言，这也说明触达不是单向投放。用户可以主动塑造自己的信息流；内容需要与对方真实需要建立联系，才能形成更稳定的观看和互动。For creators, this also shows that reach is not one-way advertising. Users can actively shape their own information feed; Content needs to connect with the other party's real needs to form more stable viewing and interaction.
11.3 长辈版指南提供了哪些具体建议11.3 What specific recommendations does the Elder Edition Guide provide?
2026年2月的原网页以长图解释推荐反馈，建议对真正喜欢的内容认真观看，再按需求点赞、评论、收藏、转发或关注作者。对不喜欢的内容则及时划过，并使用明确的负反馈工具。The original February 2026 webpage explains recommended feedback with long images, suggesting that you carefully watch content you truly like, then like, comment, save, share, or follow the author as needed. For content you don't like, scroll through promptly and use clear negative feedback tools.
指南中的“至少八成”出现在向用户说明充分观看的语境里，用来帮助表达喜好；它不是创作者必须达到的官方完播率门槛。观看某条视频的八成，与一条作品在全部观众中的完播率，也属于不同统计对象。The guide's "at least 80%" appears in contexts where users are fully watched to help express preference; It is not an official completion rate threshold that creators must meet. Eighty percent of watching a video and the completion rate of a work among all viewers are also different statistics.
评论建议围绕具体内容表达想法，而不只是重复泛泛赞美。收藏可以按用途整理，方便以后查找；关注几个确实感兴趣的作者，可以建立稳定的信息来源；主动搜索则能更直接地告诉系统当前需求。Comments suggest expressing ideas around specific content, rather than just repeating generic compliments. Favorites can be organized by purpose for easier future searches; Follow a few authors who are genuinely interested to establish a stable source of information; Active search can more directly inform the system of current needs.
这些建议形成一个简单的使用循环：主动寻找需要的内容，认真消费，留下准确反馈，保存以后会用到的作品，再通过偏好管理纠正过多或不需要的推荐。These suggestions form a simple usage cycle: actively seek out what you need, consume carefully, leave accurate feedback, save works you will use later, and then correct excessive or unnecessary recommendations through preference management.
指南还介绍了使用管理助手、AI求真和官方验证助手，把内容选择、事实辨别与服务核验结合起来。它是一份面向使用者的操作科普，不是一份账号起号或流量晋级规则。The guide also introduces the use of management assistants, AI truth-seeking, and official verification assistants, combining content selection, fact-finding, and service verification. It is a user-oriented operational guide, not an account creation or traffic promotion rules.
11.4 AI求真与官方验证助手11.4 AI Truth-Seeking and Official Verification Assistants
长辈版指南介绍了两类帮助辨别信息的功能。AI求真通过相关视频下方的提示卡、搜索等入口，帮助用户查看对争议信息的核查与解释。具体问题需要围绕原始事件、时间和对象判断，不能只看一张脱离背景的截图。The Elder Edition introduces two types of functions to help identify information. AI Truth-Seeking helps users verify and interpret disputed information through prompt cards and search entries below related videos. Specific questions need to be judged based on the original event, time, and object, not just a screenshot without context.
官方验证助手用于核验自称平台工作人员的联系信息。案例指南建议通过平台官方客服进入咨询并找到验证助手，用于核验可疑电话号码、短信或链接；其他平台的入口和支持范围应以各自官方说明为准。The official verification assistant is used to verify the contact information of a person claiming to be a platform staff member. Case guidelines recommend consulting through the platform's official customer service and finding the verification assistant to check suspicious phone numbers, messages, or links; the entry points and support scope of other platforms should follow their respective official instructions.
指南还提到让来电方提供动态验证口令，再在官方渠道核验。这里的口令由自称工作人员的一方提供，与用户把自己的短信登录验证码告诉对方是不同流程。The guidelines also mention having the caller provide a dynamic verification code, which is then verified through official channels. The code is provided by the party claiming to be a staff member, which is a different process from the user giving their own login verification code from messages.
2025年度报告网页称，官方验证助手已帮助约12万名用户。这个数字描述工具的服务触达情况。使用时应以App内相应功能实际呈现为准，网页截图反映的是发布时的界面。The 2025 annual report webpage states that the official verification assistant has helped about 120,000 users. This number describes the reach of the tool's service. Usage should be based on the actual presentation of the corresponding feature in the App; the webpage screenshot reflects the interface at the time of publication.
11.5 未成年人、长辈与使用时间管理11.5 Minors, elders, and the management of usage time
平台公开举措把时间管理、适龄内容和家庭参与放在同一个框架中。时间工具包括休息提醒、睡眠提醒、定时关闭和时间锁等；面向家庭的功能则围绕家长管理、未成年人模式及亲子协作展开。The platform's public measures place time management, age-appropriate content, and family involvement within the same framework. Time tools include break reminders, sleep reminders, timed shutdown, and time locks; family-oriented features focus on parental control, minor modes, and parent-child collaboration.
2026年举措将未成年人工作概括为健康守护、亲子护航、优质陪伴。内容供给不仅要排除不适宜内容，也需要提供不同年龄可以理解和使用的知识、兴趣与陪伴材料。The 2026 initiatives summarize work for minors as health protection, parent-child guidance, and quality companionship. Content supply should not only exclude inappropriate content but also provide knowledge, interests, and companionship materials comprehensible and usable by different ages.
2026年4月的案例网页介绍，未成年人模式内容围绕兴趣素养、知识教育、探趣百科和亲子陪伴四个主题组织，当时内容池约680万条，日均增加约6,000条知识内容。The case webpage from April 2026 introduces that content in the minor mode is organized around four themes: interest literacy, knowledge education, exploration encyclopedia, and parent-child companionship. At that time, the content pool had about 6.8 million items, with approximately 6,000 knowledge items added daily.
AI笔记用于提炼科普与课程视频的关键点，根据年龄和阅读习惯辅助呈现，并可把长视频分章，方便定位。它把推荐后的消费体验继续向学习与理解延伸。AI notes are used to extract key points from popular science and course videos, assist presentation according to age and reading habits, and can divide long videos into chapters for easy navigation. It extends the post-recommendation consumption experience toward learning and understanding.
针对长辈，网页强调清晰易懂的功能说明、偏好调整、风险信息辨别和更方便的服务入口。不同年龄的需求各不相同，产品设计因此同时考虑内容、时间与操作难度。Regarding elders, the webpage emphasizes clear and easy-to-understand feature explanations, preference adjustments, risk information identification, and more convenient service access. Needs vary by age, so product design considers content, time, and operational difficulty simultaneously.
十二、关于“窃听”和“平台强推”的公开解释12. Public Explanation of 'Eavesdropping' and 'Platform Forced Promotion'
12.1 为什么刚聊过的话题会出现在推荐或广告里12.1 Why recently discussed topics appear in recommendations or advertisements
针对“刚聊过某商品就看到广告”的疑问，原网页主要给出两条解释：已有网络行为参与广告匹配，以及人更容易记住巧合命中的情形。Regarding the question of 'seeing an ad right after chatting about a certain product,' the original webpage mainly provides two explanations: existing online behavior participating in ad matching, and people being more likely to remember coincidental hits.
程序化广告可以由广告主依据用户在其服务中的浏览、搜索、收藏等行为，向相关受众展示广告。用户在购物平台查过猫粮，之后在另一App见到该购物平台投放的猫粮广告，可以由这一机制解释，并不需要假设当前App听到了线下谈话。Programmatic advertising allows advertisers to display ads to relevant audiences based on users' browsing, search, and favorite behavior within their services. If a user has checked cat food on a shopping platform and then sees an ad for that cat food from the shopping platform on another app, this mechanism can explain it, and there is no need to assume that the current app overheard offline conversations.
此外，人每天谈论和接触很多事，大多数话题没有随后出现在信息流中，通常不会留下强烈记忆；偶然命中的一次则格外显眼。只统计命中、忽略未命中，会放大推荐与谈话之间的相关感。In addition, people discuss and encounter many things every day, most of which do not subsequently appear in their information feed and usually do not leave a strong memory; a coincidental hit stands out exceptionally. Counting only hits while ignoring misses can exaggerate the perceived correlation between recommendations and conversations.
网页同时从语音处理成本、设备权限与系统提示等角度论证平台持续监听的不合理性。对这部分内容，更准确的用途是理解平台给出的解释与假设，而不是把它扩展为对所有应用、全部设备状态的技术审计结论。The webpage also argues the unreasonableness of platforms continuously listening from the perspectives of voice processing costs, device permissions, and system prompts. For this part, the more accurate purpose is to understand the explanations and assumptions offered by the platform, rather than extending it to a technical audit conclusion for all apps and all device states.
自然内容推荐和商业广告也需要分开理解。前者主要匹配内容消费，后者还涉及广告主目标与广告投放机制；用户看到相似商品，并不能仅凭表面现象确定信号来自哪条链路。Natural content recommendations and commercial advertisements also need to be understood separately. The former mainly matches content consumption, while the latter also involves advertiser goals and ad delivery mechanisms; seeing similar products does not allow a user to determine from the surface alone which signal chain they come from.
12.2 网页中的语音成本例子怎样计算12.2 How the voice cost example on the webpage is calculated
2025年1月的网页使用当时列举的语音转写价格作假设计算：每小时0.6至1元，每人每天处理16小时，一年按365天计算。The webpage from January 2025 uses the voice transcription prices listed at the time for hypothetical calculation: 0.6 to 1 yuan per hour, 16 hours processed per person per day, calculated for 365 days a year.
如果按1亿名用户持续处理：If assuming 100 million users continuously processing:
这是一项用于解释规模成本的假设算例，不是平台真实支出，也不是2026年的服务报价。输入价格、处理时长和人数改变，结果就会同比变化。存储和后续分析等成本还未计入这个简单乘法。This is a hypothetical example used to explain scale cost, not the platform's actual expenditure, nor the service quote for 2026. If input prices, processing hours, and the number of people change, the result changes proportionally. Costs for storage and subsequent analysis are not included in this simple multiplication.
它说明高频、长期、全量处理与利用已有主动行为之间，可能存在很大的成本差异；它本身不能代替对某个具体App行为的检测。It illustrates that there could be a large cost difference between high-frequency, long-term, full-volume processing and leveraging existing active behavior; it cannot substitute for detecting a specific app's behavior.
12.3 网红走红与平台作用怎样理解12.3 How to understand influencer popularity and platform influence
关于“网红是不是被强推出来”，平台网页强调用户选择是走红的重要原因。内容能让大量用户愿意观看、点赞、分享或关注，推荐模型才有更丰富的匹配反馈；强行持续展示明显不受欢迎的内容，会带来负反馈。Regarding the question of whether "internet celebrities are forcibly promoted," the platform webpage emphasizes that user choice is an important reason for becoming popular. Content that encourages a large number of users to watch, like, share, or follow gives the recommendation model more comprehensive matching feedback; forcibly and continuously displaying content that is obviously unpopular will lead to negative feedback.
但平台作用并非只有被动读取行为。同一栏目还公开了价值目标、优质内容子链路、运营计划和高热质量评估。更完整的理解是：用户选择、算法匹配、内容供给、价值设计与平台运营共同影响传播。However, the platform's role is not only passive in reading behavior. The same section also publicly shares value goals, high-quality content sub-paths, operational plans, and high-heat quality assessments. A more complete understanding is that user choice, algorithmic matching, content supply, value design, and platform operations jointly influence dissemination.
热点事件当事人、引发广泛共鸣的单条作品或外部媒体报道，都可能带来短期关注增长。能否持续，则取决于后续内容是否继续满足受众需要，以及作者是否具备稳定经营和回应公众关注的能力。Individuals involved in hot events, single works that resonate widely, or reports from external media can all bring short-term attention growth. Whether it can continue depends on whether subsequent content continues to meet audience needs and whether the creator has the ability to maintain operations and respond to public attention.
网页还介绍，对虚假营销、低俗擦边、同质低质等问题账号，可以限制流量、涨粉或经营变现。高热度不会自动抵消内容与账号问题。The webpage also introduces that for accounts with issues such as false marketing, vulgar borderline content, or homogeneous low-quality content, traffic, follower growth, or monetization can be restricted. High popularity does not automatically offset problems with content or accounts.
对创作者更有参考价值的是区分偶发爆发与可持续增长：前者可能围绕一个事件，后者需要稳定的内容方向、受众信任和持续供给。两者的准备方式和观察指标并不相同。More valuable for creators is distinguishing between sporadic bursts and sustainable growth: the former may revolve around a single event, while the latter requires a stable content direction, audience trust, and continuous supply. The preparation methods and observation indicators for the two are not the same.
十三、平台公开工作的时间线与年度进展13. Platform Public Work Timeline and Annual Progress
13.1 从网站建设到交互式算法科普13.1 From Website Construction to Interactive Algorithm Popularization
| 时间Time |
原网页记录的节点Nodes recorded on the original webpage |
| 2025年1月3日January 3, 2025 |
提出建设安全与信任中心及10项措施，涵盖算法、治理、服务和内容生态Proposed the establishment of a Safety and Trust Center and 10 measures covering algorithms, governance, services, and content ecology |
| 2025年3月30日March 30, 2025 |
网站上线，公开算法原理、治理规则与服务机制Website launched, publicly disclosing algorithm principles, governance rules, and service mechanisms |
| 2025年4月15日April 15, 2025 |
在北京举办开放日，介绍推荐原理和治理机制，并开展交流Hosted an open day in Beijing, introducing recommendation principles and governance mechanisms, and conducted exchanges |
| 2025年5月May 2025 |
新版使用管理助手公开介绍，增强偏好可视化与自主调节The new version of the user management assistant was publicly introduced, enhancing preference visualization and independent adjustment |
| 2025年12月16日December 16, 2025 |
发布大模型推荐、RankMixer和STCA系列技术解释Release of Technical Explanations for Large Model Recommendations, RankMixer, and STCA Series |
| 2026年1月21日January 21, 2026 |
发布2025年度回顾及2026年新一轮10项举措Release of 2025 Annual Review and 10 New Initiatives for 2026 |
| 2026年3月5日March 5, 2026 |
介绍体验算法板块，使用交互动画解释推荐流程Introduction to the Experience Algorithm Section, explaining the recommendation process using interactive animations |
| 2026年3月30日March 30, 2026 |
网站周年文章报告累计访问超过300万次，文章与公告超过300篇Website anniversary article reports total visits exceeding 3 million, with over 300 articles and announcements |
| 2026年4月至6月April to June, 2026 |
发布社会价值案例、评论治理、AI求真、反网暴、多兴趣及AIGC检测文章Release of articles on social value cases, comment governance, AI truth-seeking, anti-online violence, multiple interests, and AIGC detection |
| 2026年8月24日August 24, 2026 |
发布新大众文艺报告的论坛与主要结论介绍Release of the forum and main conclusions of the new mass culture report |
网页中的访问量随时间变化：2025年度回顾为上线约九个月超过150万次，2026年3月上旬文章为超过200万次，周年文章为超过300万次。这些不是互相矛盾的同一时点统计。The page visit count over time: the 2025 Annual Review exceeded 1.5 million visits in about nine months after launch, the early March 2026 articles exceeded 2 million visits, and the anniversary article exceeded 3 million visits. These are not contradictory statistics for the same point in time.
周年文章提到的300余篇，是整个网站的文章和公告规模，并不等于本次算法栏目入口列出的27篇文章。本资料覆盖的是算法栏目及所列专题入口，另将相关页面中的治理和服务信息一并整理。The over 300 articles mentioned in the anniversary article are the total number of articles and announcements on the entire website, not equal to the 27 articles listed in the algorithm section entry. This material covers the algorithm section and the listed topic entries, and also organizes relevant governance and service information from related pages.
13.2 2025年最初公布的10项措施13.2 The 10 Measures Initially Announced in 2025
2025年1月的计划可按以下十个方向理解：The January 2025 plan can be understood in the following ten directions:
- 建设网站与线下展示，公开算法基础运作和运营机制，并提出公开热点榜单计算原理的方向。Build the website and offline exhibitions, disclose basic algorithm operations and operational mechanisms, and propose directions for publicly disclosing the calculation principles of popular ranking lists.
- 公开社区规范、审核标准与典型治理案例，增加治理投入。Disclose community norms, review standards, and typical governance cases, and increase governance investment.
- 为热点、网暴及涉企侵权相关当事人完善一站式服务。Provide one-stop services for parties involved in hot topics, online violence, and corporate infringement.
- 扶持优质内容，完善优质子链路、精选产品、作者激励与服务。Support high-quality content, improve high-quality sub-channels, curated products, author incentives, and services.
- 增强多样性、长尾探索、搜推联动、社交信号和重复过滤。Enhance diversity, long-tail exploration, search-recommendation linkage, social signals, and duplicate filtering.
- 加强谣言与虚假人设治理，推动精准辟谣和内容标注。Strengthen governance over rumors and false personas, promote precise debunking and content labeling.
- 改善评论氛围，完善防网暴工具及相关援助。Improve comment atmosphere, enhance anti-cyberbullying tools and related support.
- 优化时间管理，服务未成年人和老年人。Optimize time management, serve minors and the elderly.
- 对违法黑产加强线下协同打击。Strengthen offline coordinated crackdown on illegal activities.
- 开放意见征求、用户调研、公众交流和算法咨询渠道。Open channels for opinion solicitation, user research, public communication, and algorithm consultation.
其中，“预计增加优质内容曝光300%”属于当时的年度计划表达，不能写成每条优质作品都获得三倍增量，也不能直接作为已经完成的年度结果。年度回顾应按后来公布的实际指标另行记录。Among them, "expected increase in high-quality content exposure by 300%" was an expression of the annual plan at the time and should not be written as each piece of high-quality work gaining a threefold increase, nor should it be directly used as an already achieved annual result. Annual reviews should be recorded separately according to the actual indicators announced later.
这份计划的重要信息是平台不仅调整排序，也同时调整产品入口、内容供给、治理、服务和公众沟通。推荐体验来自这些工作的共同作用。The important information in this plan is that the platform not only adjusts ranking but also simultaneously adjusts product entry points, content supply, governance, services, and public communication. The recommendation experience comes from the combined effect of these efforts.
13.3 2026年公开的10项持续举措13.3 Ten Ongoing Initiatives Publicly Announced in 2026
2026年的安排延续公开、治理与服务，并增加对新技术和创作生态的具体表述。The 2026 arrangement continues openness, governance, and services, while adding specific statements regarding new technology and the creation ecosystem.
| 方向Direction |
主要内容Main Content |
| 算法公开Algorithm Transparency |
继续解释运作机制，介绍大模型应用与技术创新，扩大科普Continue explaining the operating mechanism, introduce applications of large models and technological innovation, expand popular science education |
| 治理透明Governance Transparency |
增加技术与资金投入，公开典型案例、标准与机制Increase technical and financial investment, disclose typical cases, standards, and mechanisms |
| 用户服务User Services |
优化95152入口、一键代办、批量及涉企侵权处理Optimize the 95152 entry point, one-click assistance, batch processing, and enterprise-related infringement handling |
| 优质生态High-Quality Ecosystem |
青年创作者、创作阶梯、AI创作浪潮、媒体精选等计划；完善优质池与收入分成Plans for young creators, creation ladders, the wave of AI creation, curated media, etc.; improve high-quality content pools and revenue sharing |
| 原创保护Original Content Protection |
优化账号健康分，治理恶意营销号并协同打击黑灰产Optimize account health scores, govern malicious marketing accounts, and coordinate the crackdown on black and gray industry |
| 可信社区Trusted Community |
治理谣言、仿冒、伪造和误导性AIGC，迭代AI求真及辟谣合作Governing rumors, counterfeits, forgeries, and misleading AIGC, iterating AI for truth-seeking and rumor debunking cooperation |
| 友善社区Friendly Community |
结合大模型治理愤怒诱饵和网暴，完善当事人保护Combining large model governance for anger bait and online harassment, improving protection for involved parties |
| 争议热点Controversial Hotspots |
强化研判，鼓励理性讨论并公开治理动作Strengthen assessment, encourage rational discussion, and publicly disclose governance actions |
| 全年龄服务All-Age Services |
完善未成年人模式、亲子护航及长辈使用指引Improve minor mode, parent-child guidance, and senior user instructions |
| 公众监督Public Supervision |
定期报告、用户反馈、公众与媒体开放日Regular reports, user feedback, public and media open days |
账号健康分在这份页面中属于持续优化的治理机制。网页没有给出一套可供创作者计算自然推荐流量的完整分值公式，因此不宜将它与推荐优先级直接等同。Account health scores in this page are part of a continuously optimized governance mechanism. The webpage does not provide a complete formula for creators to calculate natural recommendation traffic, so it should not be directly equated with recommendation priority.
13.4 2025年度网页回顾：公开、服务与优质内容13.4 2025 Annual Web Review: Transparency, Services, and Quality Content
以下按照2026年1月21日发布的年度报告原网页长图整理，保留其2025年度回顾语境。The following is organized based on the long-form webpage of the annual report released on January 21, 2026, retaining the context of the 2025 annual review.
| 主题Theme |
网页披露的进展或数量Progress or numbers disclosed on the website |
| 网站公开Website Transparency |
上线约九个月访问超过150万次Launched for about nine months, accessed over 1.5 million times |
| 算法交流Algorithm Communication |
举办30余场沟通活动，相关科普视频播放超过1,000万次Held more than 30 communication events, related science popularization videos watched over 10 million times |
| 治理公开Governance Transparency |
发布120余份治理公告，并开展规则讨论与意见交流Released over 120 governance announcements, and carried out rule discussions and opinion exchanges |
| 侵权服务Infringement Services |
同侵权内容防护、批量举报及相关处理能力建设Constructed capabilities for protecting against infringing content, batch reporting, and related handling |
| 客服Customer Service |
统一95152等入口，改进智能服务和问题办理Unified entry points like 95152, improved intelligent services and problem handling |
| 官方验证Official Verification |
验证助手帮助约12万名用户Verification assistant helped about 120,000 users |
| 风险协同Risk Coordination |
配合警方打击166个诈骗团伙，识别研判约12万条人身安全风险内容Cooperated with the police to crack down on 166 fraud gangs and identified and assessed about 120,000 pieces of personal safety risk content |
| 生命救助Life-saving |
网页报告帮助挽救1,340余人生命Web reports helped save more than 1,340 lives |
| 创作者成长Creator Growth |
青年创作者计划帮助100余位作者达到百万粉丝；媒体精选计划覆盖近百个节目The Youth Creator Program helped over 100 creators reach one million followers; the Media Selection Program covered nearly 100 programs |
| 优质内容供给High-Quality Content Supply |
优质投稿提升118%，播放提升160%High-quality submissions increased by 118%, views increased by 160% |
| 资讯内容News Content |
原创资讯播放提升60%Original news views increased by 60% |
内容项目涉及大国重器、地方体育、文化传承、非遗、出版、国家公园和城市发现等主题。它们说明平台在用户匹配之外，也通过项目组织和作者支持增加特定类型的内容供给。Content projects involve themes such as key national equipment, local sports, cultural heritage, intangible cultural heritage, publishing, national parks, and urban discoveries. They show that besides matching users, the platform also increases the supply of specific types of content through project organization and author support.
表中的提升比例沿用网页口径。供给量、播放量、参与作者数和服务人数分别回答不同问题，不能相互替代。The increase percentages in the table follow web standards. Supply volume, view volume, participating authors, and served users answer different questions and cannot replace each other.
13.5 2025年度网页回顾：治理与使用体验13.5 2025 Web Annual Review: Governance and User Experience
| 主题Theme |
年度网页列出的数据或动作Data or actions listed on the annual web page |
| 谣言治理Rumor Control |
曝光量下降90%，举报量下降74%，辟谣提醒日均约3亿次Exposure decreased by 90%, reports decreased by 74%, rumor debunking reminders averaged about 300 million per day |
| AI内容AI Content |
AI标识覆盖提升50%，治理滥用AI相关账号超过1万个AI labeling coverage increased by 50%, more than 10,000 AI-related accounts were governed for misuse |
| 网暴治理Cyberbullying Governance |
全年拦截约1.2亿条不当评论，向约132万个账号推送一键防网暴提醒，新增谁可以@我的设置Intercepted about 120 million inappropriate comments throughout the year, sent one-click anti-cyberbullying reminders to about 1.32 million accounts, added a "who can @ me" setting |
| 咨询服务Consultation Services |
2025年1月至11月为635名用户提供免费法律咨询From January to November 2025, provided free legal consultation to 635 users |
| 涉未成年人治理Minor Protection Governance |
处置约240万条违规内容、约8.4万账号次Handled about 2.4 million pieces of violating content and about 84,000 account instances |
| 色情引流Pornographic Traffic |
封禁约133万个相关账号，并协助线下抓捕66人Banned approximately 1.33 million related accounts and assisted in the offline arrest of 66 people |
| 恶意营销与黑产Malicious marketing and black industry |
治理1,000余个团伙、涉及约50万个账号，并处置相关工具Governed over 1,000 gangs, involving about 500,000 accounts, and dealt with related tools |
| 自主选择Autonomous choice |
推进使用管理助手、探索更多、内容偏好与更长历史学习Promote the use of management assistants, explore more, content preferences, and longer history learning |
年度回顾还讨论热点事件中的仿冒账号、流量炒作、事实核实和当事人保护。例如，围绕韦东奕的相关问题处置3,500个仿冒或蹭热账号；围绕武大图书馆争议邀请专业创作者梳理事实；围绕郑州续面事件向双方发送防网暴提醒，处理地域攻击及谣言等内容。The annual review also discusses impersonation accounts, traffic manipulation, fact verification, and protection of parties involved in hot events. For example, 3,500 impersonation or trending accounts were dealt with regarding issues around Wei Dongyi; professional creators were invited to clarify facts regarding the Wuhan University Library controversy; both sides of the Zhengzhou re-examination incident were sent online violence prevention reminders, and regional attacks and rumors were handled.
这些内容补充了推荐算法之外的运行环境：有些内容被减少传播，可能与真实性、账号行为或对当事人的影响相关；有些用户体验改善来自服务工具，而不是排序模型本身。This content supplements the operational environment beyond recommendation algorithms: some content has reduced exposure, possibly due to authenticity, account behavior, or impact on the parties involved; some user experience improvements come from service tools rather than the ranking model itself.
十四、原网页中的创作与社会价值案例14. Cases of Creation and Social Value on the Original Webpage
14.1 日常生活、知识讲解与工程现场14.1 Daily life, knowledge explanation, and engineering site
2026年4月的案例页面先展示三种不同的内容价值。The case page from April 2026 first presents three different types of content value.
李福贵记录在乡村售卖豆腐、点心和与村民相处的日常，形成围绕乡愁、真实生活与情感共鸣的交流。页面报告，截至2025年底其粉丝超过1,000万，视频总播放约22亿次，并用这一案例解释留言与作者回复形成的握手互动。Li Fugui records the daily life of selling tofu and snacks in the countryside and interacting with villagers, forming exchanges around nostalgia, real life, and emotional resonance. The page reports that, by the end of 2025, he had over 10 million fans, with a total video view count of about 2.2 billion, and uses this case to explain the handshake interaction formed by comments and author replies.
混子哥边画边讲通过白板手绘解释古代官制，《丞相&宰相》视频播放超过1.2亿次，点赞约324万。页面借此说明专业用户的评论、收藏等深度互动，可以帮助识别知识内容的价值。Hunzi Ge explains the ancient official system while drawing on a whiteboard; the video 'Chancellor & Prime Minister' has over 120 million views and about 3.24 million likes. The page uses this to illustrate how deep interactions such as professional user comments and collections can help identify the value of knowledge content.
小Lin说进入海底高铁隧道施工现场，把大型工程转化为公众可以理解的内容。相关铁路建设话题登上热榜第一，页面称触达超过6,800万网友。这个案例同时涉及真实现场、创作者表达和大国重器计划的资源支持。Xiao Lin enters the construction site of the undersea high-speed railway tunnel and translates large-scale engineering into content understandable by the public. Related railway construction topics reached first on the trending list, and the page states that it reached over 68 million netizens. This case also involves the real site, creator expression, and resource support from the national major equipment plan.
三个案例对应三条不同路径：真实生活形成共鸣，清楚讲解降低理解门槛，现场访问提供稀缺信息。其共同点是内容本身提供了用户愿意消费和讨论的理由，不能仅用一个互动率解释全部结果。The three cases correspond to three different paths: resonance with real life, clear explanations lowering the barrier to understanding, and on-site visits providing scarce information. What they have in common is that the content itself provides users with reasons to be willing to consume and discuss; it cannot explain all results by a single engagement rate.
14.2 深度长视频与非遗表演14.2 In-depth Long Videos and Intangible Cultural Heritage Performances
风雪夜灯以约16.8万字文稿梳理254年的战国历史，相关超长视频播放超过1亿，单条作品带来约100万粉丝增长。页面用这一案例说明，评论、收藏、分享等多目标可以帮助深度长内容找到受众。Wind and Snow Night Light uses a manuscript of about 168,000 words to trace 254 years of Warring States history, with related ultra-long videos playing over 100 million views, and a single work generating about 1 million new followers. The page uses this case to illustrate that comments, favorites, sharing, and other objectives can help deep long-form content find its audience.
同一网页报告，2025年超过30分钟的中长视频数量同比增长207%，累计时长达到3.2亿分钟。这里的累计时长用于描述页面所述内容规模，不能自动换算为用户实际观看时长。The same webpage report shows that by 2025, the number of medium-to-long videos over 30 minutes will increase by 207% year-on-year, with a cumulative total time reaching 320 million minutes. The cumulative time here is used to describe the content scale described on the page and cannot be automatically converted to actual viewing time by users.
蔡冬梅长期学习提线木偶戏，2021年开始在案例平台直播，平均每年上百场，积累近180万粉丝，并获得更多演出展示机会。页面把收藏、复访和追更等长期目标与非遗寻找新观众联系起来。Cai Dongmei has long studied string puppetry and began livestreaming on the case platform in 2021, averaging over a hundred performances annually, accumulating nearly 1.8 million followers and gaining more performance opportunities. The page links long-term goals such as collecting, revisiting, and following up with intangible cultural heritage seeking new audiences.
对深度内容，结构与解释帮助用户持续理解；对表演类内容，技艺特色与持续演出帮助建立稳定关系。不同类型的长期价值，可以通过不同的行为组合被观察到。For in-depth content, structure and explanation help users maintain continuous understanding; For performance content, technical features and ongoing performances help establish stable relationships. Different types of long-term value can be observed through various behavioral combinations.
这些成功案例适合用来理解机制和内容特点。它们展示的是已经发生的结果，不能用作所有新作者的平均增长预期。These success stories are suitable for understanding mechanisms and content characteristics. They show results that have already occurred and cannot be used as an average growth expectation for all new authors.
14.3 从内容兴趣连接到消费、学习与澄清14.3 Connecting content interests to consumption, learning, and clarification
2025年苏超相关内容播放超过370亿次、分享超过1.1亿次。平台围绕赛事设置本地生活活动和商家团购链接，将观看、搜索、消费与核销连接起来。网页报告赛事期间江苏团购订单量同比增长24%，销售额增长68%。By 2025, Suzhou Super League-related content will exceed 37 billion views and more than 110 million shares. The platform has set up local lifestyle activities and merchant group-buying links around the tournament, linking viewing, searching, consumption, and redemption. Web Report During the tournament, Jiangsu's group-buying orders increased by 24% year-on-year, and sales revenue rose by 68%.
搜推联动在这里连接主动需求与推荐发现：有人先看赛事内容产生兴趣，再搜索城市、美食或出行信息；系统也可以利用搜索表达的需求补充后续内容与服务推荐。同期消费增长涉及赛事热度、活动和供给等多种因素，不能仅归因于某一个模型。Search and Push Linkage connects proactive needs and recommendations here, discovering: some people first become interested after watching the event content, then search for city, food, or travel information; The system can also use the search to express needs to supplement subsequent content and service recommendations. Contemporaneous consumption growth involves multiple factors such as event heat, activities, and supply, and cannot be attributed to a single model.
学习案例通过未成年人模式与AI笔记，把知识视频转化为重点提炼和章节访问，帮助用户更好消费已经推荐到的内容。Learning cases transform knowledge videos into key highlights and chapter access through juvenile modes and AI notes, helping users better consume content that has already been recommended to them.
辟谣案例则把事实核查与搜索卡片、官方澄清视频、定向推荐结合起来，让看过错误信息的人更有机会接触后续纠正。Debunking cases combine fact-checking with search cards, official clarification videos, and targeted recommendations, giving people who have seen misinformation a better chance to encounter subsequent corrections.
这三类场景说明推荐的价值不止是多一次播放，还包括满足具体消费需求、提升学习效率和修正信息误解。These three types of scenarios illustrate that the value of recommendations is not just an additional play but also includes meeting specific consumption needs, improving learning efficiency, and correcting misunderstandings.
14.4 新大众文艺的十大领域与四个价值方向14.4 Ten Major Areas and Four Value Directions of New Popular Arts
2026年8月的报告发布网页介绍了算法与普通人创作之间的双向关系：推荐帮助作品找到受众，丰富的创作和真实反馈也帮助系统学习什么内容有价值。The webpage introducing the report released in August 2026 explains the bidirectional relationship between algorithms and ordinary creators: recommendations help works find an audience, while rich creations and genuine feedback help the system learn what content is valuable.
网页列出的十个领域包括：素人才艺、微短剧、网络文学、文化演艺、生活万象、大众音乐创作、非遗与民间手艺、大众体育与全民健身、素人写作、知识素养。The ten areas listed on the webpage include: amateur talent, micro short dramas, online literature, cultural performances, everyday life, popular music creation, intangible cultural heritage and folk crafts, mass sports and national fitness, amateur writing, and knowledge literacy.
这些领域并非都依靠同一种消费方式。音乐与表演可能产生重复观看和分享，知识内容可能形成收藏与复访，系列剧情可能带来关注与追更，生活记录可能形成持续的情感连接。多目标体系为不同内容价值提供了不同的识别入口。These areas do not all rely on the same mode of consumption. Music and performances may lead to repeat views and sharing, knowledge content may result in collection and revisits, serial plots may bring attention and following, and life records may create continuous emotional connections. A multi-objective system provides different entry points for recognizing different content values.
发布页以ECHO模型概括四个方向：更有活力、更有文化、更加美好、更加开放。网页没有展开该缩写每个字母的英文释义，本资料沿用其已给出的中文维度。The release page summarizes the four directions with the ECHO model: more energetic, more cultural, better, and more open. The webpage does not elaborate the English meanings of each letter in the abbreviation; this material follows the Chinese dimensions provided.
这一页所讨论的公共文化空间，重点在于普通人的才艺、职业经验、生活记录与情感表达能够进入更广泛的传播。算法参与机会分配，但作品的构思、表达和文化价值仍来自人的创造。The public cultural space discussed on this page focuses on allowing ordinary people's talents, professional experience, life records, and emotional expression to reach wider dissemination. Algorithms participate in opportunity allocation, but the conception, expression, and cultural value of works still come from human creation.
14.5 课文短片、早餐店写作与AI影像14.5 Text Short Films, Breakfast Shop Writing, and AI Imaging
郝家豁团队以真人实拍把《孔乙己》《金色的鱼钩》等课文转化为短片。团队并非影视科班出身，作品通过观众反馈不断改进，并被部分教师用于课堂。其价值来自对文本的理解、场景重建与影像表达。The Haojiahuo team transforms texts like "Kong Yiji" and "Golden Fishhook" into short films with real-life shooting. The team is not formally trained in film, and their works are continuously improved through audience feedback and have been used by some teachers in classrooms. Their value comes from understanding the text, reconstructing scenes, and visual expression.
江南好在湖北经营早餐店，中年重新拾起写作，自2019年以来在案例平台写下50多万字。职业生活提供了经验与观察，网络发表和读者回应则为持续表达提供了空间。Jiangnan runs a breakfast shop in Hubei and picked up writing again in middle age. Since 2019, he has written more than 500,000 words on a case platform. Professional life provides experience and observation, while online publication and reader feedback offer space for continuous expression.
编导李让从舞蹈专业转向自学AIGC影像创作，代表作《告别》播放超过1亿次。发布页强调，生成工具可以参与画面制作，但故事、分镜、构思与情感组织仍需要创作者完成。Director Li Rang switched from a dance major to self-learning AIGC image creation. His representative work "Farewell" has been played over 100 million times. The release page emphasizes that generative tools can participate in image production, but the story, storyboards, concept, and emotional organization still need to be completed by the creator.
这三个案例对应低成本参与创作的不同方式：已有文学文本的再表达、日常经验的写作、技术工具辅助的影像。制作门槛下降后，更值得区分的是内容有没有独特观察、完整表达和打动受众的理由。These three cases correspond to different ways of low-cost participation in creation: re-expression of existing literary texts, writing about daily experiences, and image creation assisted by technical tools. After the production threshold is lowered, what is more worth distinguishing is whether the content has unique observation, complete expression, and a reason to move the audience.
AI工具与内容质量不是同义关系。工具能缩短某些制作步骤，作品是否有价值仍要看创意和最终体验；是否让用户误认制作方式或真实事件，则是另一个需要清晰交代的问题。AI tools and content quality are not synonymous. Tools can shorten certain production steps, but whether a work has value still depends on creativity and the final experience; whether it causes users to misunderstand the production method or real events is another issue that needs to be clearly explained.
十五、把公开机制用于创作：怎样提高推荐机会15. Using Public Mechanisms for Creation: How to Improve Recommendation Opportunities
15.1 一条更容易持续分发的内容，需要同时解决什么15.1 What must a piece of content that is easier to continuously distribute address simultaneously
从公开机制可以推导出五个创作问题：谁需要这条内容，开头能否让对方识别价值，主体是否兑现承诺，作品是否值得保存或分享，扩大传播后是否仍经得起质量判断。这是一套内容分析方法，不是平台公布的操作公式。From public mechanisms, five creative questions can be deduced: who needs this content, can the beginning let the audience recognize its value, does the main body deliver on its promise, is the work worth saving or sharing, and after expanding distribution, does it still withstand quality judgment. This is a set of content analysis methods, not an operational formula published by the platform.
首先，明确接收者。与其只写“做美食”，不如明确“给第一次做饭的人解释怎样判断肉熟了”。具体需求有助于选题、结构和例子保持一致，也让用户更快判断是否值得继续看。First, clarify the recipient. Rather than just writing "make food," it is better to specify "explain to someone cooking for the first time how to tell if meat is cooked." Specific needs help keep topics, structure, and examples consistent, and also let users quickly judge whether it is worth continuing to watch.
其次，把价值交代清楚。开头可以说明问题、呈现结果或展示值得理解的反常现象；主体要提供支持这个承诺的信息。开头吸引与内容兑现之间的落差，会影响实际观看和后续信任。Second, make the value clear. The beginning can state a problem, present a result, or show an anomaly worth understanding; the main body should provide information supporting this promise. The gap between an attractive beginning and the content’s fulfillment affects actual viewing and subsequent trust.
再次，选择适合内容的价值目标。操作教程可以追求方便复用，解释型内容可以追求理解与讨论，故事可以追求共鸣与分享，系列创作可以追求持续关注。内容定位不同，理想反馈组合也不同。Third, choose an appropriate value goal for the content. Tutorials can pursue ease of reuse, explanatory content can pursue understanding and discussion, stories can pursue resonance and sharing, and series creations can pursue continuous attention. Different content positioning leads to different ideal feedback combinations.
最后，让信息、素材与表达经得起更广泛人群的理解。流量扩大后，作品面对的受众更复杂，原本依赖小圈子背景的玩笑或省略条件，可能被不同方式理解。补足必要语境，同时保持表达清晰，有助于提高传播的稳定性。Finally, ensure that information, materials, and expressions can be understood by a broader audience. As traffic increases, the audience becomes more diverse, and jokes or assumptions that originally relied on small-circle contexts may be interpreted in different ways. Providing necessary context while keeping expressions clear helps improve the stability of communication.
15.2 不同内容类型应该优先改进什么15.2 What should be prioritized for improvement in different content types
| 内容类型Content Type |
用户主要需要Primary user needs |
优先改进Priority improvements |
适合观察的反馈Feedback to observe |
| 教程与工具Tutorials and Tools |
按步骤完成任务Completing tasks step by step |
条件、步骤、参数、常见错误、结果Conditions, steps, parameters, common mistakes, results |
收藏、复访、具体问题Favorites, revisits, specific questions |
| 知识解释Knowledge Explanation |
理清复杂问题Clarifying complex issues |
结构、例子、因果、术语说明Structure, examples, cause and effect, terminology explanation |
有效观看、分享、深度评论Effective viewing, sharing, in-depth comments |
| 生活记录Life Documentation |
真实感与共鸣Authenticity and resonance |
具体细节、人物关系、情绪节奏Specific details, relationships between people, emotional rhythm |
分享、评论交流、持续关注Sharing, commenting, continuous following |
| 技艺与表演Skills and Performance |
审美与技艺体验Aesthetic and skill experience |
关键动作、完整呈现、特色Key actions, complete presentation, features |
重复观看、关注、追更Repeated viewing, following, updating |
| 系列内容Series Content |
连续解决一组问题Solving a series of problems continuously |
单集价值、系列递进、稳定主题Single episode value, series progression, consistent theme |
关注后观看、后续访问Watching after following, subsequent visits |
| 资讯与现场News and On-Site Reporting |
获得新信息Obtaining new information |
时间地点、事实依据、背景解释Time and place, factual basis, background explanation |
有效讨论、分享、信息纠正情况Effective discussion, sharing, information correction |
| AI影像AI Imaging |
新表达与故事体验New Expressions and Story Experiences |
构思、情感、叙事、制作方式说明Explanation of Concept, Emotion, Narrative, and Production Method |
观看质量、分享、对作品的讨论Viewing Quality, Sharing, and Discussion of Works |
这张表用于选择改进方向，并不意味着系统只为每类内容计算表中的目标。多目标仍同时存在，只是不同类型的内容更容易在不同维度体现价值。This table is used to select improvement directions and does not mean the system calculates only the goals in the table for each type of content. Multiple goals still exist simultaneously, but different types of content more easily reflect value in different dimensions.
所谓优化收藏，不只是增加一句“记得收藏”，而是让用户以后回来确实能找到需要的信息；优化关注，也不只是请求关注，而是让用户理解以后还能获得什么。Optimizing collections is not just adding a line like 'remember to save,' but ensuring users can actually find the information they need when they return; optimizing followers is not just asking for follows, but helping users understand what they can gain in the future.
15.3 创作者可以自己计算哪些指标15.3 Creators Can Calculate Their Own Metrics
下面是内容复盘的通用计算方法。分子和分母应采用同一统计周期，并与后台定义保持一致；人数、次数和播放量不要混用。Below is a general method for content review calculations. The numerator and denominator should use the same statistical period and be consistent with backend definitions; do not mix people, counts, and play volumes.
| 指标Metric |
一种明确的计算口径A Clear Calculation Standard |
用途Purpose |
| 完播率Completion Rate |
完整播放次数÷播放开始次数Full Plays ÷ Play Starts |
观察完成消费的比例Observe the proportion of complete consumption |
| 平均观看时长Average Watch Duration |
总观看时长÷对应播放次数Total Watch Time ÷ Corresponding Plays |
观察实际投入时间Observe actual time invested |
| 平均观看比例Average Watch Ratio |
平均观看时长÷视频时长Average Watch Duration ÷ Video Length |
辅助比较相近形式内容Assist in comparing similar content formats |
| 点赞率Like Rate |
点赞数÷选定口径的播放量Number of Likes ÷ Play Volume of Selected Standard |
观察即时认可Observe immediate approval |
| 收藏率Collection Rate |
收藏数÷同口径播放量Number of Collections ÷ Same Standard Play Volume |
观察保存意愿Observe willingness to save |
| 分享率Share Rate |
分享数÷同口径播放量Number of Shares ÷ Same Standard Play Volume |
观察转发意愿Observe willingness to share |
| 关注转化率Pay attention to conversion rate |
可归因于作品的新增关注÷同口径播放量New follows attributable to the work ÷ plays measured by the same criteria |
观察持续接收意愿Observe willingness to continue receiving |
| 评论回复覆盖率Comment reply coverage rate |
获得作者回复的评论数÷纳入统计的评论数Number of comments that received an author reply ÷ number of comments included in the statistics |
观察作者回应情况Observe the author's response situation |
例如，一条60秒视频有1,000次播放开始，350次完整播放，总观看45,000秒，60个赞、40次收藏、10个可归因新增关注，则按上述示例口径：完播率35%，平均观看45秒，平均观看比例75%，点赞率6%，收藏率4%，关注转化率1%。For example, a 60-second video starts with 1,000 plays, 350 complete plays, total watch time 45,000 seconds, 60 likes, 40 saves, 10 new follows attributable to the work; according to the example criteria above: completion rate 35%, average watch 45 seconds, average watch proportion 75%, like rate 6%, save rate 4%, follow conversion rate 1%.
平均观看比例与完播率不同。平均看45秒，不等于75%的人看完。重复播放还可能使某些口径下的平均观看比例超过100%。如果后台使用有效播放、去重人数或其他定义，应按其口径重新计算，不能直接混比。Average watch proportion is different from completion rate. Watching an average of 45 seconds does not mean 75% of people watched the whole video. Repeat plays may also cause the average watch proportion to exceed 100% under certain criteria. If the backend uses valid plays, deduplicated users, or other definitions, it should be recalculated according to that standard and not directly compared across different criteria.
这些历史指标可用于复盘，但并不等于推荐模型对下一个特定用户预测的概率。These historical metrics can be used for review, but they do not equate to the recommendation model's probability prediction for the next specific user.
15.4 怎样比较两条作品才有意义15.4 How to meaningfully compare two works
比较作品时，尽量先固定内容类型、时长范围、目标人群和观察时段。一个十秒笑点与一个二十分钟教程，直接比完播率，往往无法回答哪条更满足用户需求。When comparing works, try to first fix content type, duration range, target audience, and observation period. Comparing the completion rate of a 10-second comedic clip with a 20-minute tutorial often cannot answer which one better meets user needs.
可以围绕一个明确假设做小规模实验。例如，保持同类选题与表达长度，比较“先交代使用场景”和“先展示最终结果”两种开头，观察早期流失、平均观看与收藏是否同步变化。每次尽量只改一个主要因素，减少解释上的混乱。Small-scale experiments can be done around a clear hypothesis. For example, keeping the same topic and expression length, compare the two openings “show the use scenario first” and “show the final result first,” and observe whether early drop-offs, average watch, and saves change in sync. Each time, try to change only one main factor to reduce explanatory confusion.
数据越少，偶然波动越大。100次播放中的5个赞与6个赞，差一个人的动作就变化一个百分点；当流量扩大到不同受众时，指标下降也可能反映人群变化。先观察多条同类作品，再形成稳定判断，比凭单条结果给账号定性更可靠。The smaller the data set, the greater the random fluctuations. In 100 plays, 5 likes versus 6 likes changes the rate by one percent just from one person's action; when traffic expands to different audiences, metric declines may also reflect audience changes. Observing multiple works of the same type first and then forming stable judgments is more reliable than characterizing an account based on a single result.
复盘最好同时看过程与结果。过程包括开头是否清楚、哪一段难懂、评论反复问什么；结果包括观看、保存、分享和关注。用户提出的具体疑问，往往比一个总分更能指向下一次该修改哪里。It's best to review both the process and the results. The process includes whether the beginning is clear, which parts are difficult to understand, and what questions are repeatedly asked in the comments. The results include views, saves, shares, and follows. Specific questions raised by users often point to where adjustments should be made next more effectively than an overall score.
可以为每次实验记录：目标受众、用户问题、主要价值、改动因素、统计窗口、关键结果与下一步动作。这里的统计窗口由自身业务需要选择，不存在由本栏目公布的统一发布后小时数。You can record for each experiment: target audience, user questions, main value, factors of change, statistical window, key results, and next steps. The statistical window here should be chosen based on your own business needs; there is no uniform number of hours after release published by this column.
15.5 播放量停下来时，按什么顺序检查15.5 When playback stops, check in what order
先检查作品状态和平台是否有具体通知。明确的内容处置、审核提示或账号问题，比仅凭曲线猜测更直接。没有这些证据时，再检查匹配与消费过程。First, check the work status and whether the platform has specific notifications. Explicit content handling, review prompts, or account issues are more direct than guessing based only on the curve. If there is no such evidence, then check the matching and consumption process.
| 观察到的现象Observed phenomena |
可以检查的问题Questions to check |
更有用的改进方向More useful directions for improvement |
| 很多人迅速划过Many people quickly scroll past |
开头能否让目标用户识别价值，画面和信息是否清晰Can the beginning let the target users recognize the value, and are the visuals and information clear |
更早交代问题或结果，减少无关铺垫Introduce the problem or result earlier, reduce irrelevant padding |
| 开头能留人，中途流失明显The beginning can retain viewers, but mid-video drop-off is obvious |
承诺是否兑现，结构是否断裂，解释是否过难Are promises fulfilled, is the structure broken, is the explanation too difficult |
重组叙述，增加必要例子，压缩重复Reorganize the narrative, add necessary examples, compress repetition |
| 观看较好，保存较少Views are good, saves are few |
作品是否具有以后再用的价值Does the work have value for later use |
对适用题材补足步骤、参数和索引Supplement steps, parameters, and indexes for applicable topics |
| 单条互动好，关注较少Single interactions are good, follows are few |
用户是否知道作者以后能持续提供什么Do users know what the author can continue to provide in the future |
建立清晰主题与连续问题体系Establish a clear theme and a continuous problem system |
| 评论很多但攻击性强Many comments but highly aggressive |
讨论是否偏离内容价值，表达是否制造无意义对立Does the discussion deviate from content value, is expression creating meaningless opposition |
重新组织问题，回应事实与经验Reorganize questions, respond to facts and experiences |
| 早期表现好，扩大后下降Early performance is good, declines after expansion |
新受众是否需要同样背景知识Do new audiences require the same background knowledge? |
补足语境，检查受众定位Supplement context, check audience targeting |
| 同类作品连续下降Similar works decline consecutively |
重复感、需求变化、题材供给是否变化Repetition, changing demand, has the supply of themes changed? |
研究真实新问题，调整内容而非只换包装Research real new problems, adjust content rather than just changing the packaging |
这张表提供排查线索。流量停滞本身不是某个原因的确证，也不应直接推出“系统在等待付费”“到了固定流量池”或“账号永久失去推荐”等结论。This table provides clues for investigation. Traffic stagnation itself is not proof of a particular reason, nor should it directly lead to conclusions such as 'the system is waiting for payment,' 'it has reached a fixed traffic pool,' or 'the account has permanently lost recommendations.'
15.6 最初两篇文章，究竟藏着哪些热门线索15.6 The first two articles, what popular clues are really hidden?
最初提供的第一篇链接是《平台治理为推荐算法设置“护栏”》（文章15384）。它与热门直接相关的内容有三点：内容传播越广，经历的评估越多；更高播放量会触发更严格的复审与质量判断；潜在高热内容阶段重视原创和优质，运营人员可能介入综合研判。The first link provided initially is "Platform Governance Sets 'Guardrails' for Recommendation Algorithms" (Article 15384). There are three points in its content directly related to popularity: the wider the content spreads, the more evaluations it undergoes; higher view counts trigger stricter re-examination and quality assessment; in the potential high-traffic content phase, originality and quality are emphasized, and operation personnel may intervene in comprehensive judgments.
因此，这篇文章给出的线索是“持续扩大传播还需要满足什么条件”。它没有公布每个互动动作的流量兑换值，也没有给出一张固定播放量台阶表。Therefore, the clue from this article is 'what conditions need to be met for continued widespread dissemination.' It does not disclose the traffic value of each interaction, nor does it provide a fixed view count step chart.
第二篇《从零开始了解推荐系统》（文章15358）解释的是用户与内容怎样匹配，包括协同过滤、向量表示、深度学习和双塔召回。它帮助理解“作品为什么会先被挑给某些用户”，也说明标签名称不足以概括实际匹配关系。The second article, "Understanding Recommendation Systems from Scratch" (Article 15358), explains how users and content are matched, including collaborative filtering, vector representation, deep learning, and dual-tower recall. It helps understand 'why works are first shown to certain users' and also shows that label names are insufficient to summarize the actual matching relationships.
把这两篇与多目标、优质子链路和评论质量文章放在一起，才更完整：内容先要匹配到潜在受众，再在多种价值上形成有效反馈，同时经过质量与治理判断，才可能持续扩大推荐。Putting these two together with articles on multi-objectives, high-quality sub-paths, and comment quality gives a more complete picture: content must first match the potential audience, then generate effective feedback across multiple value dimensions, while also passing quality and governance assessments, before it can continue to expand recommendations.
所以，关于热门最值得研究的不是一个隐藏常数，而是受众、内容价值、长期关系和传播适宜性如何共同作用。Therefore, when it comes to popularity, what is most worth studying is not a hidden constant, but how audience, content value, long-term relationships, and dissemination suitability interact.
十六、专题页面与学习路线16. Topic Pages and Learning Paths
16.1 看懂算法：十集课程目录16.1 Understanding the Algorithm: Ten-Episode Course Catalog
“看懂算法”专题提供十集课程，目录与时长如下。这里整理网页目录和简介，不将未观看的视频内容当作文字稿补写。The "Understanding Algorithms" series offers ten courses. The catalog and durations are as follows. Here, the webpage catalog and descriptions are organized, and unwatched video content is not supplemented as transcripts.
| 集数Episode |
课程名称Course Name |
时长Duration |
| 1 |
视频推荐流背后的隐形机器The Invisible Machine Behind Video Recommendation Streams |
09:14 |
| 2 |
算法如何把用户和内容变成数据How Algorithms Turn Users and Content into Data |
09:32 |
| 3 |
在海量内容里打捞你的兴趣Fishing for Your Interests in Massive Content |
09:55 |
| 4 |
推荐系统如何兼顾速度与精度How Recommendation Systems Balance Speed and Accuracy |
09:09 |
| 5 |
推荐结果最终如何产生How Recommendation Results Are Ultimately Generated |
09:11 |
| 6 |
推荐系统的价值罗盘The Value Compass of Recommendation Systems |
10:05 |
| 7 |
推荐系统如何面对陌生用户和新内容How Recommendation Systems Handle New Users and New Content |
08:43 |
| 8 |
大模型如何为下一代推荐系统赋能How Large Models Empower the Next Generation of Recommendation Systems |
08:15 |
| 9 |
如何在多元价值评估体系中寻找平衡How to Find Balance in Multi-Value Evaluation Systems |
06:04 |
| 10 |
平台如何守住内容安全边界How Platforms Maintain Content Safety Boundaries |
05:48 |
此外，专题还设置算法机制解析、算法多目标体系、算法思维成长等科普分类及相关活动入口。课程适合配合本文从匹配、排序、价值目标到治理依次学习。In addition, the series also provides popular science categories and related activity entry points on algorithm mechanism analysis, multi-objective algorithm systems, and algorithm thinking growth. The courses are suitable to be studied in sequence along with this article, from matching, ranking, and value goals to governance.
16.2 体验算法：从召回到多样性的互动演示16.2 Experiencing Algorithms: Interactive Demonstration from Recall to Diversity
体验算法页面通过引子、推荐系统、双塔召回、兴趣时钟、Wide & Deep排序、随机扰动和结语组织流程。The experience algorithm page is organized through an introduction, recommendation system, dual-tower recall, interest clock, Wide & Deep ranking, random perturbation, and conclusion.
双塔部分展示用户与内容表示怎样用于匹配；兴趣时钟让用户观察时段变化与内容偏好的关系；排序部分解释记忆已知关联与泛化潜在关联；随机扰动部分展示相关性与内容多样性的调整。The dual-tower section shows how user and content representations are used for matching; the interest clock lets users observe the relationship between time periods and content preferences; the ranking section explains remembering known associations and generalizing potential associations; the random perturbation section demonstrates adjustments to relevance and content diversity.
这些交互有助于把抽象模型转化为可感知过程，但演示只选取推荐链路中的部分步骤。虚拟角色、推荐指数、时段与扰动强度属于教学设置，不能用来反推真实账号的阈值或参数。These interactions help transform abstract models into perceptible processes, but the demonstration only selects certain steps in the recommendation pipeline. Virtual characters, recommendation indexes, time periods, and perturbation intensities are teaching settings and cannot be used to infer real account thresholds or parameters.
本次整理读取了页面说明与可见流程，排序子模块曾停留在加载状态，因此未把其未呈现的交互结果作为证据。与该入口配套的2026年3月文章，提供了召回、排序和多样性环节的文字解释。This compilation read page descriptions and visible flows; the sorting submodule was once loaded and therefore did not present interactions as evidence. The accompanying March 2026 article provides textual explanations of recalls, sorting, and diversity steps.
16.3 按目的选择阅读顺序16.3 Choose the reading order according to the purpose
想理解热门分发，可以先读多目标公式与优质内容机制，再读高热审核、评论质量、长尾兴趣，最后使用创作复盘章节。这样能同时看到匹配、价值与传播条件。To understand popular distribution, first read the multi-objective formula and quality content mechanism, then focus on high-heat review, review quality, long-tail interest, and finally use the creation review section. This way, you can see matching, value, and dissemination conditions simultaneously.
想理解技术实现，可以从协同过滤、向量与双塔开始，再读特征交叉、行为预测、STCA、Trinity和RankMixer。注意这些技术分别作用于表示、候选寻找、历史理解与排序效率。To understand technical implementations, start with collaborative filtering, vectors and dual towers, then read feature crossing, behavior prediction, STCA, Trinity, and RankMixer. Note that these techniques are applied respectively to representation, candidate search, historical understanding, and sorting efficiency.
想改善自己的使用体验，可以先看使用管理助手、不感兴趣、主动搜索、收藏与关注，再了解多样性和长期兴趣机制。To improve your user experience, you can first check out Usage Management Assistant, Disinterest, Active Search, Favorites & Follows, and then understand the mechanisms of diversity and long-term interest.
想了解平台治理，可以按照标准制定、发布审核、评论语境、谣言检索、AIGC识别、网暴风险研判的顺序阅读，再用年度举措和案例理解这些能力怎样落地。To understand platform governance, you can read in the order of standard formulation, release review, comment context, rumor retrieval, AIGC identification, and cyberbullying risk assessment, then use annual initiatives and case studies to understand how these capabilities are implemented.
信息公开入口把基础文章、最近更新和专题页面放在一起。按问题组织阅读，比只按文章发表顺序阅读，更容易建立完整的理解。The information disclosure portal brings together basic articles, recent updates, and special topic pages. Organizing reading by question is easier to build a complete understanding than reading only in the order articles were published.
十七、常见问题速查17. Quick Reference for Frequently Asked Questions
17.1 热门、分数与流量池17.1 Popular, Score, and Traffic Pool
新媒体平台有没有公开推荐公式？ 本文引用的平台案例公开了简化逻辑与多目标乘法表达，展示行为预测、价值权重和目标融合；完整函数、参数及各场景配置没有一并公开。Do social media platforms publicly release recommendation formulas? The platform cases cited in this article publicly express simplified logic and multi-objective multiplication, demonstrating behavior prediction, value weighting, and goal integration; complete functions, parameters, and scenario configurations are not disclosed together.
一个赞、一个收藏分别等于多少分？ 本栏目没有给出统一兑换值。模型计算的是特定用户对候选内容的预测与价值，而不是把后台动作数直接按固定单价相加。How many points do one like and one favorite each have? This section does not provide a unified exchange value. The model calculates the prediction and value of candidate content by specific users, rather than directly adding up the backend action count at a fixed unit price.
完播率达到多少就一定上热门？ 没有适用于所有内容的公开门槛。时长、受众、分享、收藏、关注和长期反馈都会影响评价。What is the completion rate to guarantee going viral? There is no public threshold for all content. Duration, audience, sharing, favorites, followers, and long-term feedback all affect reviews.
初始流量一定是500吗？ 原网页提到基础流量，但没有公布人人统一的500次初始分发规则。Is the initial traffic always 500? The original webpage mentioned basic traffic, but it did not publish a unified rule of 500 initial distributions for everyone.
复审与三审就是固定流量池吗？ 不是同一个概念。原文描述治理节点，并明确播放量阈值波动；实际还可被举报、质疑和异常信号触发。Is re-review and three-review a fixed traffic pool? Not the same concept. The original text describes governance nodes and clearly states that the view count threshold fluctuates; in practice, it can also be triggered by reports, disputes, and abnormal signals.
自然推荐和热榜是一套公式吗？ 本栏目没有提供足以证明两者相同的完整公式。个性化推荐针对用户与内容匹配，榜单涉及特定范围内的热点排序，应分别理解。Are natural recommendations and trending charts the same formula? This section does not provide a complete formula sufficient to prove that the two are the same. Personalized recommendations are aimed at matching users with content, while charts involve ranking of hot topics within a specific range, and should be understood separately.
17.2 内容质量与互动17.2 Content Quality and Interaction
为什么收藏值得关注？ 收藏表达保存意愿，后续复访进一步体现持续使用价值，有助于识别高信息价值内容。Why is saving to favorites worth attention? Saving expresses the intention to keep content, and subsequent revisits further reflect continuous use value, which helps identify content with high informational value.
回复每条评论都能获得额外流量吗？ 握手模型关注留言得到作者回应的积极互动，网页没有公布逐条回复与流量的固定关系。Does replying to every comment gain additional traffic? The handshake model focuses on positive interactions when the author responds to messages; the webpage did not announce a fixed relationship between each reply and traffic.
争议越大是不是越容易火？ 评论质量与恶评也进入目标，依靠攻击和引战产生的互动可能限制分发。Does greater controversy make content more likely to go viral? The quality of comments and even negative comments also count toward the target; interaction generated solely through attacks or provocation may limit distribution.
好评论只能是赞美吗？ 不是。客观信息、合理建议、具体问题、经验分享和有理由的不同观点都可能有交流价值。Do good comments have to be praise? No. Objective information, reasonable suggestions, specific issues, experience sharing, and well-reasoned differing viewpoints can all have communication value.
懂行用户就是认证账号吗？ 页面没有把两者等同。其核心是相关领域反馈对内容质量的识别价值，而不是一个公开的认证加分表。Is a knowledgeable user the same as a verified account? The page does not equate the two. Its core is the recognition value of feedback from relevant fields on content quality, not a public verification points system.
长视频是不是天然吃亏？ 多目标体系引入分享、收藏、关注、时长等目标，帮助识别完播率之外的价值。长视频仍需要适合长表达的内容和结构。Are long videos inherently at a disadvantage? Multi-objective systems include goals such as sharing, saving to favorites, following, and watch time to help identify value beyond completion rate. Long videos still need content and structure suitable for longer expression.
17.3 兴趣、标签与模型17.3 Interests, Tags, and Models
账号标签越少越精准吗？ 固定文字标签不足以解释实际推荐。网页介绍的向量、神经网络和聚类可以表达更复杂的用户—内容关系。Are fewer account tags more precise? Fixed text tags are insufficient to explain actual recommendations. Vectors, neural networks, and clustering introduced on the webpage can express more complex user-content relationships.
为什么突然看一类视频后，推荐变得重复？ 近期行为可能形成强信号，同时系统也通过多样性、长期兴趣和用户控制功能平衡重复。Why does the recommendation become repetitive after suddenly watching a certain type of video? Recent behavior may form a strong signal, while the system also balances repetition through diversity, long-term interests, and user control features.
做冷门题材一定更容易得到扶持吗？ 长尾机制关注用户确实喜欢且分发不足的兴趣，冷门本身不是充分条件。Is it easier to get support for niche topics? The long-tail mechanism focuses on interests that users truly like but are under-distributed; being niche alone is not a sufficient condition.
以前喜欢但很久没看的内容，系统还能记住吗？ Trinity使用较长周期兴趣统计和历史种子，STCA探索更充分利用长行为序列，两者从不同层面支持长期记忆。Can the system still remember content I liked but haven't watched for a long time? Trinity uses long-cycle interest statistics and historical seeds, while STCA explores making fuller use of long behavioral sequences; both support long-term memory from different perspectives.
Trinity、STCA、RankMixer是一回事吗？ 分别重点处理兴趣组织、超长历史和排序模型计算效率，可以互相补充。Are Trinity, STCA, and RankMixer the same thing? They respectively focus on interest organization, ultra-long history, and ranking model computing efficiency, and can complement each other.
探索更多会清空历史偏好吗？ 页面描述的是增加多样性和调整权重，后续浏览习惯仍然影响推荐。Will exploring more content erase historical preferences? The page describes increasing diversity and adjusting weights; subsequent browsing habits still influence recommendations.
17.4 AI、治理与使用疑问17.4 AI, governance, and usage questions
使用AI制作就会被限制推荐吗？ 页面支持优质AI创作，强调清晰标识，并治理误导、仿冒和违规内容；AI制作方式本身不能直接等同于低质量。Will content created with AI be restricted in recommendations? The page supports high-quality AI creations, emphasizes clear labeling, and governs misleading, counterfeit, and violating content; the use of AI itself cannot be directly equated with low quality.
检测AI只看手指和画面瑕疵吗？ 还包括频域特征、音频、口型、多模态信号、水印和标注等。Does AI detection only look at fingers and image flaws? It also includes frequency domain features, audio, lip movements, multimodal signals, watermarks, and annotations.
AI求真只是让大模型凭记忆判断吗？ 它结合资料库、网络检索、多轮查证、信源分级与必要的人工复核。Does AI Truth-Seeking only let the large model make judgments based on memory? It combines databases, web searches, multi-round verification, source grading, and necessary human review.
“算法不懂内容”与“大模型理解语义”矛盾吗？ 早期文章强调行为关联无需人的语义理解，后续文章讨论新增语义能力与不同模型。需要按技术对象和发表时间理解。Is "algorithm doesn't understand content" contradictory to "large models understand semantics"? Early articles emphasized behavior correlation without human semantic understanding, while later articles discuss additional semantic capabilities with different models. It should be understood according to the technical object and publication time.
长辈指南的八成观看是起号门槛吗？ 不是。那是向观看者说明充分观看怎样表达喜好，与作品总体完播率门槛不同。Is 80% view rate for elder guides a threshold for starting an account? No. That explains to viewers how full watching expresses preference and is different from the overall completion rate threshold for content.
看到刚聊过的商品就能证明App在窃听吗？ 单次命中不能确定信号来源。网页提供了已有行为参与广告匹配和巧合被放大的解释。Does seeing a recently discussed product prove the App is eavesdropping? A single occurrence cannot determine the signal source. The page provides explanations involving prior behavior in ad matching and amplified coincidences.
十八、术语与公式速查18. Quick Reference for Terms and Formulas
18.1 推荐系统常用术语18.1 Common terms in recommendation systems
| 术语Term |
在本文中的含义Meaning in this article |
| 候选内容Candidate content |
尚未最终展示、等待匹配或排序的内容Content not yet finally displayed, waiting for matching or ranking |
| 召回Recall |
从海量内容中快速找出可能相关的候选Quickly find potentially relevant candidates from massive content |
| 粗排Coarse ranking |
用较低成本进一步缩小候选范围Further narrow down the candidate range at a lower cost |
| 精排Fine ranking |
对较少候选进行更细致的行为与价值预估Conduct more detailed behavior and value estimation for fewer candidates |
| 混排或重排Mixed ranking or re-ranking |
结合多样性、重复控制等组织最终列表Organize the final list by combining diversity, repetition control, etc. |
| 冷启动Cold start |
新用户或新内容历史反馈不足时的匹配问题Matching problem when new users or new content lack historical feedback |
| 协同过滤Collaborative filtering |
利用不同用户的共同选择发现偏好关联Discover preference correlations using the common choices of different users |
| 矩阵分解Matrix factorization |
用较低维度表示近似用户与内容关系Represent the user-content relationship approximately in a lower dimension |
| Embedding |
将对象或特征转换为可计算的向量表示Convert objects or features into computational vector representations |
| 双塔Dual tower |
分别编码用户与内容，再计算匹配关系的结构Structure that encodes users and content separately, then calculates the matching relationship |
| 余弦相似度Cosine similarity |
比较向量方向接近程度的度量A measure comparing the closeness of vector directions |
| 特征交叉Feature crossing |
学习多个条件组合出现时的意义Learn the significance when multiple condition combinations occur |
| 泛化Generalization |
将已有学习用于没有完全重复出现过的情形Apply existing learning to situations that have not completely repeated |
| 行为预估Behavior prediction |
预测用户可能发生的交互或消费行为Predict interactions or consumption behaviors that users may perform |
| 多目标Multi-objective |
同时考虑多种用户、作者与生态价值Consider multiple user, author, and ecosystem values simultaneously |
| 目标权重Objective weight |
表达某目标在融合中的相对作用Express the relative effect of an objective in the fusion |
| 长期组合目标Long-term combined objective |
如收藏后复访、关注后追更等跨时间行为关系Cross-time behavior relationships such as revisiting after collecting or following after subscribing |
| 内容簇Content Cluster |
根据相似关系形成的内容集合A collection of content formed based on similarity relationships |
| 长尾兴趣Long-tail Interests |
小众、低频但可能稳定存在的偏好Niche, low-frequency but potentially stable preferences |
| 兴趣直方图Interest Histogram |
用户对不同内容簇的长期行为分布The long-term behavior distribution of users across different content clusters |
| Token |
模型处理的信息单元，可表示特征而非文字The unit of information processed by a model, which can represent features rather than text |
| 自注意力Self-Attention |
序列内部不同位置之间的信息匹配与聚合Information matching and aggregation between different positions within a sequence |
| 交叉注意力Cross-Attention |
一组查询从另一组信息中提取相关部分A set of queries extracting relevant parts from another set of information |
| FFN |
对表示进行变换和提炼的前馈网络Feedforward Network for transforming and refining representations |
| MoE |
通过多个专家模块组合或选择性计算的结构A structure combining or selectively computing through multiple expert modules |
| MFU |
衡量模型对硬件计算能力有效利用程度的指标A metric that measures how effectively a model utilizes hardware computational capacity |
18.2 内容治理与AI常用术语18.2 Content Governance and Common AI Terms
| 术语Term |
在本文中的含义Meaning in this article |
| 初审Preliminary Review |
上传后的初步风险识别与判断Initial risk identification and judgment after uploading |
| 复审、三审Re-examination, third review |
在后续传播阶段进行的进一步评估Further assessment conducted during the subsequent dissemination stage |
| 多模态Multimodal |
同时利用文字、图像、声音、视频等信息Simultaneously using text, images, audio, video, and other information |
| OCR |
从图片等视觉材料中识别文字Recognizing text from images or other visual materials |
| ASR |
把语音转成文字的自动语音识别Automatic speech recognition that converts speech into text |
| Agent |
能围绕任务组织分析、工具使用与后续步骤的智能系统Intelligent systems capable of organizing analysis, tool usage, and subsequent steps around a task |
| RAG |
检索相关材料，再辅助模型组织回答或判断Retrieving relevant materials and assisting the model in organizing answers or judgments |
| 信源分级Source Grading |
按来源属性与可信程度组织证据优先级Organizing evidence priorities according to source attributes and credibility |
| 谣言变体Rumor Variants |
改换表述或部分背景后重新传播的相似误导信息Similar misleading information that is spread again after changing the expression or some background |
| 正向评论目标Positive Comment Goals |
预测并鼓励具有交流价值的评论Predict and encourage comments with communicative value |
| 恶评目标Target negative comments |
预测并抑制攻击、引战等不良评论Predict and suppress harmful comments such as attacks and trolling |
| RLHF |
利用人类反馈及奖励信号优化模型Optimize the model using human feedback and reward signals |
| DPO |
利用成对偏好样本直接优化模型Directly optimize the model using paired preference samples |
| AIGC |
由人工智能生成的内容Content generated by artificial intelligence |
| 显式水印Explicit watermark |
用户能够直接看到或听到的生成标记Generation marks that users can directly see or hear |
| 隐式水印Implicit watermark |
需借助相应识别方法读取的标记Marks that need to be read using corresponding recognition methods |
| 空间域Spatial domain |
图像像素在画面位置上的表示Representation of image pixels at positions in the frame |
| 频域Frequency domain |
图像变化按频率组织的表示Representation of image variations organized by frequency |
| FFT |
快速傅里叶变换，用于频域分析Fast Fourier Transform, used for frequency domain analysis |
| 伪影Artifacts |
生成或处理过程中形成的异常视觉痕迹Abnormal visual traces formed during generation or processing |
| 困难样本Difficult samples |
难以靠简单特征区分、用于提高学习能力的样本Samples that are difficult to distinguish using simple features, used to improve learning ability |
| 存量回扫Stock sweep |
用更新的治理能力重新检查已有内容Recheck existing content with updated governance capabilities |
18.3 所有公式分别解决什么问题18.3 What problems do all formulas solve
| 表达Expression |
用途Purpose |
本资料中的性质Properties in this material |
| 行为预测与价值权重结合Behavior prediction combined with value weighting |
概括推荐优先级Summarize recommendation priority |
原网页公开简式Original web page public simplified version |
| 各目标经过函数处理后相乘Targets are multiplied after function processing |
展示多目标融合结构Display multi-objective fusion structure |
原网页公式图转写Transcription of original web page formula diagrams |
| 加权和与乘法数值例子Weighted sum and multiplication numerical examples |
演示权重和变换如何影响排序Demonstrate how weights and transformations affect ranking |
本文自设教学例子This article sets up its own teaching examples |
| 用户向量与内容向量点积User vectors and content vector dot product |
解释低维关系预测Explain low-dimensional relationship prediction |
通用数学说明General mathematical explanation |
| 余弦相似度Cosine similarity |
解释向量方向相似程度Explain the similarity of vector directions |
通用数学说明General mathematical explanation |
| 注意力公式Attention formula |
解释查询、键、值的信息聚合Explain the aggregation of information in queries, keys, and values |
通用数学说明General mathematical explanation |
| 平方级与线性复杂度Quadratic and linear complexity |
解释长历史计算的增长方式Explain how long-history computation grows |
对技术页面的数学说明Mathematical explanation for technical pages |
| 完播、平均观看、收藏等比率Completion, average watch time, favorites, and other ratios |
帮助创作者复盘已发生的数据Help creators review past data |
本文明确口径的分析方法Analytical methods with clearly defined terms in this article |
| 时价×时长×天数×人数Current price × duration × days × number of people |
说明网页的假设语音处理成本Explain the assumed speech processing cost of the webpage |
对网页算例的展开Elaboration of webpage calculation examples |
推荐分数、相似度、行为概率、历史互动率和复杂度不是同一个量。分清它们的用途，才能判断一个公式是否真的能够回答手头的问题。Recommendation scores, similarity, behavior probability, historical interaction rate, and complexity are not the same quantity. Only by distinguishing their purposes can you judge whether a formula truly answers the question at hand.
参考文献References
- S01 新媒体平台安全与信任中心：从零开始了解推荐系统。2025-03-30；访问日期：2026-09-04。S01 Social Media Platform Safety and Trust Center: Understanding Recommendation Systems from Scratch. 2025-03-30; Accessed: 2026-09-04.
- S02 新媒体平台安全与信任中心：用户行为背后的算法推荐逻辑。2025-01-20；访问日期：2026-09-04。S02 Social Media Platform Safety and Trust Center: Algorithmic Recommendation Logic Behind User Behavior. 2025-01-20; Accessed: 2026-09-04.
- S03 新媒体平台安全与信任中心：新媒体平台算法的多目标平衡。2025-01-20；访问日期：2026-09-04。S03 Social Media Platform Safety and Trust Center: Multi-Objective Balance in the Social Media Platform Algorithm. 2025-01-20; Accessed: 2026-09-04.
- S04 新媒体平台安全与信任中心：平台治理为推荐算法设置“护栏”。2025-01-20；访问日期：2026-09-04。S04 Social Media Platform Safety and Trust Center: Platform Governance Sets "Guardrails" for the Recommendation Algorithm. 2025-01-20; Accessed: 2026-09-04.
- S05 新媒体平台安全与信任中心：清华联合新媒体平台发布「新大众文艺」报告：“算法向善”激发全民创造。2026-08-24；访问日期：2026-09-04。S05 TikTok Safety and Trust Center: Tsinghua University and TikTok jointly release the "New Popular Arts" report: 'Algorithm for Good' inspires nationwide creativity. 2026-08-24; Accessed on: 2026-09-04.
- S06 新媒体平台安全与信任中心：新媒体平台AI反网暴行动，一场AI协同的治理探索。2026-06-12；访问日期：2026-09-04。S06 TikTok Safety and Trust Center: TikTok AI Anti-Cyberbullying Campaign, an exploration of AI collaborative governance. 2026-06-12; Accessed on: 2026-09-04.
- S07 新媒体平台安全与信任中心：打破“信息茧房”：新媒体平台推荐算法的多兴趣建模。2026-06-15；访问日期：2026-09-04。S07 TikTok Safety and Trust Center: Breaking the 'Information Cocoon': Multi-interest Modeling of TikTok's Recommendation Algorithm. 2026-06-15; Accessed on: 2026-09-04.
- S08 新媒体平台安全与信任中心：AI内容真假难辨，新媒体平台算法如何守护社区真实。2026-06-15；访问日期：2026-09-04。S08 TikTok Safety and Trust Center: AI Content Hard to Discern True from False, How TikTok's Algorithm Protects Community Authenticity. 2026-06-15; Accessed on: 2026-09-04.
- S09 新媒体平台安全与信任中心：谣言止于技术：一文读懂新媒体平台AI求真大模型工作原理。2026-05-21；访问日期：2026-09-04。S09 TikTok Safety and Trust Center: Rumors Stop at Technology: Understanding the Working Principles of TikTok AI Truth-Seeking Large Model in One Article. 2026-05-21; Accessed on: 2026-09-04.
- S10 新媒体平台安全与信任中心：新媒体平台如何用大模型治理评论区“引战”。2026-05-11；访问日期：2026-09-04。S10 TikTok Safety and Trust Center: How TikTok Uses Large Models to Govern 'Flame-Baiting' in Comment Sections. 2026-05-11; Accessed on: 2026-09-04.
- S11 新媒体平台安全与信任中心：《新媒体平台算法社会价值报告》发布：算法连接美好的8个故事。2026-04-27；访问日期：2026-09-04。S11 TikTok Safety and Trust Center: Release of the "TikTok Algorithm Social Value Report": 8 Stories Where the Algorithm Connects Happiness. 2026-04-27; Accessed on: 2026-09-04.
- S12 新媒体平台安全与信任中心：新媒体平台安全与信任中心网站上线一周年，让信任可见、让安全可感。2026-03-30；访问日期：2026-09-04。S12 TikTok Safety and Trust Center: TikTok Safety and Trust Center Website Celebrates One-Year Launch, Making Trust Visible and Safety Tangible. 2026-03-30; Accessed on: 2026-09-04.
- S13 新媒体平台安全与信任中心：新媒体平台持续加强算法透明、平台治理和用户服务，提升安全与信任。2026-01-21；访问日期：2026-09-04。S13 TikTok Safety and Trust Center: TikTok Continues to Strengthen Algorithm Transparency, Platform Governance, and User Services, Enhancing Safety and Trust. 2026-01-21; Accessed on: 2026-09-04.
- S14 新媒体平台安全与信任中心：《2025新媒体平台安全与信任年度报告》发布：全年推进算法公开，谣言曝光量下降90%。2026-01-21；访问日期：2026-09-04。S14 TikTok Safety and Trust Center: Release of the '2025 TikTok Safety and Trust Annual Report': Promoted Algorithm Transparency Throughout the Year, Rumor Exposure Decreased by 90%. 2026-01-21; Accessed on: 2026-09-04.
- S15 新媒体平台安全与信任中心：新媒体平台安全与信任中心上线“体验算法”板块：一条视频如何被推荐。2026-03-05；访问日期：2026-09-04。S15 Social Media Platform Safety and Trust Center: Social Media Platform Safety and Trust Center launches the 'Experience Algorithm' section: How a video gets recommended. 2026-03-05; Accessed: 2026-09-04.
- S16 新媒体平台安全与信任中心：新春将至，一份爸妈拿起来就能看得懂的“算法调教指南”请查收。2026-02-11；访问日期：2026-09-04。S16 Social Media Platform Safety and Trust Center: As the Spring Festival approaches, here is an 'Algorithm Training Guide' that parents can understand at a glance. 2026-02-11; Accessed: 2026-09-04.
- S17 新媒体平台安全与信任中心：大模型重塑下一代推荐系统：新媒体平台的探索。2025-12-16；访问日期：2026-09-04。S17 Social Media Platform Safety and Trust Center: Large models reshaping the next-generation recommendation system: Social Media Platform’s exploration. 2025-12-16; Accessed: 2026-09-04.
- S18 新媒体平台安全与信任中心：用算法驱动算力：新媒体平台如何打造更高效的推荐大模型。2025-12-16；访问日期：2026-09-04。S18 Social Media Platform Safety and Trust Center: Driving computing power with algorithms: How Social Media Platform builds more efficient recommendation large models. 2025-12-16; Accessed: 2026-09-04.
- S19 新媒体平台安全与信任中心：个性化推荐的新突破：新媒体平台解锁用户行为学习上限。2025-12-16；访问日期：2026-09-04。S19 Social Media Platform Safety and Trust Center: A new breakthrough in personalized recommendations: Social Media Platform unlocks the upper limit of user behavior learning. 2025-12-16; Accessed: 2026-09-04.
- S20 新媒体平台安全与信任中心：新媒体平台将建立安全与信任中心，推进算法和平台治理透明化。2025-01-03；访问日期：2026-09-04。S20 Social Media Platform Safety and Trust Center: Social Media Platform will establish a Safety and Trust Center to promote transparency in algorithms and platform governance. 2025-01-03; Accessed: 2026-09-04.
- S21 新媒体平台安全与信任中心：一文全解！新媒体平台算法原理公开。2025-04-15；访问日期：2026-09-04。S21 Social Media Platform Safety and Trust Center: Full explanation! Social Media Platform algorithm principles disclosed. 2025-04-15; Accessed: 2026-09-04.
- S22 新媒体平台安全与信任中心：新媒体平台推出新版“使用管理助手”。2025-05-21；访问日期：2026-09-04。S22 Social Media Platform Safety and Trust Center: Social Media Platform launches a new version of the 'Usage Management Assistant'. 2025-05-21; Accessed: 2026-09-04.
- S23 新媒体平台安全与信任中心：践行算法向善，新媒体平台多措并举赋能用户便捷管理推荐内容。2025-05-20；访问日期：2026-09-04。S23 Social Media Platform Safety and Trust Center: Practicing algorithmic goodness, Social Media Platform takes multiple measures to empower users to conveniently manage recommended content. 2025-05-20; Accessed: 2026-09-04.
- S24 新媒体平台安全与信任中心：算法和用户都想打破“信息茧房”。2025-01-20；访问日期：2026-09-04。S24 Social Media Platform Safety and Trust Center: Both algorithms and users want to break the 'information cocoon'. 2025-01-20; Accessed: 2026-09-04.
- S25 新媒体平台安全与信任中心：网红是平台“强推”出来的吗。2025-01-20；访问日期：2026-09-04。S25 Social Media Platform Safety and Trust Center: Are internet celebrities 'strongly promoted' by the platform? 2025-01-20; Accessed: 2026-09-04.
- S26 新媒体平台安全与信任中心：App会“窃听”用户谈话吗。2025-01-20；访问日期：2026-09-04。S26 TikTok Safety and Trust Center: Does the app 'eavesdrop' on users' conversations? 2025-01-20; Accessed: 2026-09-04.
- S27 新媒体平台安全与信任中心：新媒体平台为何突出“收藏”按钮。2025-01-20；访问日期：2026-09-04。S27 TikTok Safety and Trust Center: Why TikTok highlights the 'Favorite' button. 2025-01-20; Accessed: 2026-09-04.
- S28 新媒体平台安全与信任中心：看懂算法。网页未标注发布日期；访问日期：2026-09-04。S28 TikTok Safety and Trust Center: Understanding the algorithm. Webpage not marked with publication date; Accessed: 2026-09-04.
- S29 新媒体平台安全与信任中心：体验算法：一条视频如何被推荐。网页未标注发布日期；访问日期：2026-09-04。S29 TikTok Safety and Trust Center: Experiencing the algorithm: how a video is recommended. Webpage not marked with publication date; Accessed: 2026-09-04.
- S30 新媒体平台安全与信任中心：新媒体平台安全与信任中心：信息公开。网页未标注发布日期；访问日期：2026-09-04。S30 TikTok Safety and Trust Center: TikTok Safety and Trust Center: Information disclosure. Webpage not marked with publication date; Accessed: 2026-09-04.