【作者】 刘奇；

【导师】 连德富；

【作者基本信息】 中国科学技术大学 ， 计算机应用技术， 2025， 博士

【摘要】 在数字技术日益普及的背景下,全球数据总量呈爆炸式增长,信息过载挑战随之加剧。推荐系统作为应对这一挑战的核心技术,凭借个性化的信息匹配机制,提升了信息分发的效率与质量。为应对海量数据与严苛时延的双重约束,工业级推荐系统普遍采用由“召回—粗排—精排—重排”构成的多阶段漏斗式架构。然而,推荐系统中的用户交互数据并非源自信息全集的无偏随机抽样,而是建立在系统历史推荐与曝光基础之上的。这一闭环反馈机制引发了选择偏差。多阶段架构虽然提升了计算效率,但在层层筛选与传递的过程中,会放大选择偏差,并沿因果链条逐步衍生,最终演化为一系列相互关联的系统性偏差。具体而言,选择偏差首先在模型知识的滞后性与信息池的持续动态更新耦合下,引发了时间偏差,导致新物品无法曝光。其次,选择偏差与时间偏差的长期累积固化了马太效应,诱发了流行度偏差,致使海量长尾物品因交互数据稀疏而面临表征学习困境。最终,上述三种偏差还会在交互层面进一步演化为多反馈偏差,即系统对点击等稠密反馈的强化显著强于对购买等稀疏反馈的学习,损害了多任务推荐模型的整体性能。这一系列偏差相互交织、层层递进,不仅削弱了模型性能与用户体验,更对平台生态的公平性、多样性与可持续发展构成了威胁。本文以工业级推荐系统的偏差为中心,针对偏差诊断、时间偏差、流行度偏差和多反馈偏差四个研究对象,构建了一个贯穿诊断到治理全过程的研究框架,从解决偏差诊断的基础性难题入手,延伸至对各偏差的治理,开展了如下研究:1.面向偏差诊断的数据集构建。当前针对偏差问题的研究普遍局限于曝光空间采集所得的推荐数据集,此类数据因缺失召回、粗排等早期阶段的过滤信息,无法准确揭示偏差在多阶段系统内部的生成与传播机理。为弥合此项研究空白,本文依托于短视频平台快手实际业务场景中的工业级推荐系统,构建并发布了业内首个全链路推荐数据集。该数据集完整地记录了物品从全量候选库历经多阶段漏斗筛选,最终被曝光并接收用户多种类型反馈的全流程日志,为剖析系统性偏差的演化路径提供了实证基础,并为相关算法的研究提供了真实闭环的离线评估基准。2.面向时间偏差的物品表征在线索引算法设计。在新物品持续涌入的动态环境中,召回阶段所依赖的物品表征索引需保持及时更新,否则将导致新物品无法被召回。全量物品索引重构因计算成本高昂,难以满足工业级推荐系统的高频更新需求。现有的在线索引更新算法存在检索效果随时间迅速劣化的问题,亦无法适用于工业场景。本文提出了一种高效的在线索引更新算法,首次将性能优异的加性量化算法拓展至动态数据场景。借助矩阵的增量更新理论与随机块束搜索算法,该方法能够在不依赖历史数据的前提下,基于新物品对索引实时增量更新,从而缓解了时间偏差问题。该方法具有当前该问题最优的遗憾界理论保障。3.面向流行度偏差的长尾物品表征优化算法设计。流行度偏差使得长尾物品的交互数据稀疏,导致其表征学习的监督信号不足,难以准确建模用户对其偏好。工业推荐系统为节省训练资源,采用的负样本降采样策略,加剧了这一问题造成的负面影响。现有方法存在表征优化效果差、算法复杂难部署和因果假设性太强等问题。本文设计了一种轻量级的对比匹配辅助学习任务。通过构建两个基于对比学习的辅助监督任务,驱动模型学习用户与物品、以及物品与物品之间更具判别力的深层语义关联,尤其为样本稀疏的长尾物品引入了有效的训练信号。该方法提升了长尾物品的表征质量与推荐准确率,减轻流行度偏差,且具备在工业环境中便捷部署的优势,在美团广告业务提升了长尾物品6.37%的点击率。4.面向多反馈偏差的任务表征解耦模型设计。为精准刻画用户的多维偏好,推荐系统需协同优化点击、收藏、购买等多元目标。然而,用户的多种反馈稀疏度差异悬殊且背后意图复杂,导致当前的多任务推荐模型因难以对任务表征进行解耦而产生负迁移效应。现有方法专注于改进参数共享的机制或梯度融合策略,忽视了共享底层表征退化问题。本文提出了一种新颖的多任务推荐模型,突破了当前的底层共享-顶层分离模型结构。该架构引入超网络为各个任务动态生成专属的行为序列编码器,并辅以任务独立的特征精炼模块,从模型的最低层实现了不同任务表征学习的解耦。这缓解了多反馈偏差导致的任务间梯度冲突与表征混淆,提升了对用户多维偏好的综合预估能力,在阿里巴巴短视频广告业务提升了1%的点击率。更多还原

【Abstract】 With the increasing prevalence of digital technologies,the global volume of data has been experiencing explosive growth,thereby intensifying the challenge of information overload.As a core technology to address this challenge,recommender systems improve the efficiency and quality of information distribution through personalized information-matching mechanisms.To cope with the dual constraints of massive data and stringent latency requirements,industrial-grade recommender systems commonly adopt a multi-stage funnel architecture consisting of recall,coarse ranking,fine ranking,and re-ranking.However,user interaction data in recommender systems are not obtained through unbiased random sampling from the entire information space,but are instead generated based on the system’s historical recommendations and exposures.This closed-loop feedback mechanism gives rise to selection bias.While the multistage architecture improves computational efficiency,it also amplifies existing selection bias during successive filtering and propagation,which progressively evolves along the causal chain into a series of interrelated systemic biases.Specifically,selection bias,when coupled with the lag in model knowledge and the continuous dynamic updates of the item pool,first gives rise to temporal bias,resulting in the inability of new items to gain exposure.Second,the long-term accumulation of selection bias and temporal bias reinforces the Matthew effect,thereby inducing popularity bias and leaving massive long-tail items with sparse interactions that hinder effective representation learning.Finally,these three biases further evolve at the interaction level into multi-feedback bias,whereby the system disproportionately reinforces dense feedback such as clicks while under-learning from sparse feedback such as purchases,thus impairing the overall performance of multi-task recommendation models.Intertwined and progressively escalating,this series of biases not only undermines model performance and user experience but also poses potential threats to the fairness,diversity,and sustainability of the platform ecosystem.Therefore,this dissertation centers on biases in industrial-grade recommender systems,focusing on four research objects—bias diagnosis,temporal bias,popularity bias,and multi-feedback bias—and establishes a comprehensive research framework spanning the entire process from diagnosis to mitigation,starting with the fundamental challenge of bias diagnosis and extending to the governance of each specific bias through the following studies:1.Dataset Construction for Bias Diagnosis.Current research on bias is generally limited to recommendation datasets collected from the final exposure stage.Such datasets lack filtering information from earlier stages such as recall and coarse ranking,and therefore cannot accurately reveal the mechanisms by which bias is generated and propagated within multi-stage systems.To fill this research gap,this dissertation leverages the industrial-grade recommender system of the short-video application Kuaishou to construct and release the first full-chain recommendation dataset in the industry.This dataset comprehensively records the entire process in which items pass through multi-stage funnel filtering from the full candidate pool,are eventually exposed,and receive multiple types of user feedback.It provides an empirical foundation for analyzing the evolutionary paths of systemic biases and establishes a realistic closed-loop offline benchmark for related algorithmic research.2.Online Indexing Algorithm Design for Temporal Bias.In a dynamic environment where new items continuously emerge,the item representation index relied upon in the recall stage must be updated in a timely manner;otherwise,new items cannot be effectively recalled.However,full reconstruction schemes are prohibitively expensive in terms of computational cost and thus fail to meet the high-frequency update requirements of industrial-grade recommender systems.Existing online index update algorithms suffer from rapid degradation of recall effectiveness over time,making them unsuitable for industrial scenarios.To address this challenge,this dissertation proposes an efficient online index update algorithm that,for the first time,extends the high-performing additive quantization algorithm to streaming data scenarios.By leveraging matrix incremental update theory and a randomized block beam search algorithm,the method achieves real-time,incremental index updates for new items without relying on historical data,thereby effectively mitigating temporal bias.Moreover,the proposed approach is equipped with the optimal regret bound guarantee for this problem to date.3.Representation Optimization Algorithm for Long-Tail Items to Mitigate Popularity Bias.Popularity bias results in extremely sparse interaction data for longtail items,leading to insufficient supervisory signals in their representation learning process and making it difficult for models to accurately capture user preferences.Moreover,to reduce training costs,industrial recommender systems commonly adopt negative sampling strategies,which further exacerbate the adverse effects of this problem.Existing approaches suffer from limited representation optimization effectiveness,algorithmic complexity that hinders deployment,and overly strong causal assumptions.To address this challenge,this dissertation designs a lightweight auxiliary contrastive matching task.By constructing two auxiliary supervision tasks based on contrastive learning,the proposed framework drives the model to learn more discriminative deep semantic associations between users and items as well as between items and items,thereby introducing effective training signals particularly for long-tail items with sparse data.This method significantly improves the representation quality and recommendation accuracy of long-tail items,mitigating popularity bias,thus offering practical advantages for deployment in industrial environments.In the industrial deployment at Meituan’s advertising platform,the click-through rate of long-tail items increased by 6.37%.4.Task Representation Decoupling Model for Multi-Feedback Bias.To accurately capture users’ multi-dimensional preferences,recommender systems must jointly optimize multiple objectives such as clicks,favorites,and purchases.However,the substantial sparsity differences among these feedback types and the complexity of their underlying intentions cause existing multi-task recommendation models to suffer from negative transfer,as they struggle to decouple task-specific representations.Current methods primarily focus on improving parameter-sharing mechanisms or gradientfusion strategies,while overlooking the degradation of shared bottom representations.To address this issue,this dissertation proposes a novel multi-task recommendation model that breaks through the conventional bottom-layer sharing and top-layer separation paradigm.The proposed architecture employs a hypernetwork to dynamically generate task-specific behavior sequence encoders and incorporates task-independent feature refinement modules,thereby achieving representation decoupling across tasks from the lowest model layers.This approach effectively alleviates gradient conflicts and representation entanglement induced by multi-feedback bias,significantly enhancing the system’s ability to comprehensively estimate users’ multi-dimensional preferences.The proposed approach improves the click-through rate by 1%in Alibaba’s short video advertising business.更多还原