为应对信息过载问题，满足用户个性化的信息需求，推荐系统已经被广泛应用于多种互联网场景，深度参与用户的信息获取和决策过程。推荐系统旨在从行为数据中学习用户个性化的偏好，发现并推荐符合其兴趣的信息以提升用户的满意度。然而，用户行为具有复杂性且易受到各种偏置影响，用户满意度具有主观性和多元性。这些挑战极大地限制了推荐系统的效果。因此，对用户行为和满意度更全面的研究，建模用户的交互行为模式及背后的主观认知过程，设计面向用户多元化满意度需求的评价方法，成为支撑推荐系统发展的重点方向，也是领域学术研究的前沿热点问题。本文围绕用户行为、认知过程、满意度三个核心要素开展系统性的深入研究，主要工作和创新成果如下：(1) 针对已有推荐系统用户行为模型对质量偏置及用户停止行为研究不足的问 题，结合用户实验及真实推荐系统的大规模用户交互日志分析，探究了信息质量对用户行为的偏置效应，进而提出模型显著优化了隐式偏好反馈的构建效果；建模了用户会话停止行为的原因及模式，提出了对停止意图的识别与提前预测方法，在真实数据集上相比基线模型识别准确性提升了20.3%；(2) 针对现有系统建模中对复杂动态的用户偏好刻画不足的问题，创新性地基于 动态多阶段过程研究了用户偏好的感知及变化规律，提出对用户实际偏好的估计方法，相比基线方法准确率提升10.6%；建模了广泛存在但隐式的负向体验，提出了用户负向体验识别方法，精确度达到77.3%；(3) 针对目前推荐系统评价方法与用户满意度低一致性的问题，提出基于实际偏好估计的在线评价修正指标，相比常用在线指标，与满意度的相关性提升20.1%；提出基于外部偏好标注的新型离线评价方法，显著减轻了离线与在线评价之间的差异，开拓了新的系统评价思路；(4) 针对推荐系统的多维度评价，建模了用户对推荐解释的感知过程，提出了对 解释效果的离线评价方法，并通过与满意度的对比实验验证了其有效性；提出了系统不公平性的溯源分析和风险诊断方法，在多领域数据上进行了验证。 本文以用户为核心，深入建模了行为、认知及满意度，提出并解决了推荐系统多方面的重要挑战与问题，全面优化了系统的构建与评价，对个性化推荐系统的用户满意度研究和应用发展有着重要的推动意义。
To address the information overload problem and meet users' personalized information needs, recommender systems have been widely used in various internet scenarios, deeply participating in users' information acquisition and decision-making processes. Recommender systems aim to learn users' personalized preferences from behavioral data and retrieve information that matches users' interests to improve their satisfaction. However, user behavior is complex and biased to various factors, and user satisfaction is subjective and multifaceted. These challenges greatly limit the effectiveness of recommender systems. Therefore, modeling users' interaction behavior patterns and subjective cognitive processes, and designing better training and evaluation methods, have become key directions to support the development of recommender systems, and are also frontier in academic research. This thesis conducts systematic and in-depth research around the three key elements: user behavior, cognitive process, and satisfaction. The main work and innovations are as follows: (1) To address the ignorance of quality bias and user stopping behavior in existing user behavior models of recommender models, combining user experiments and analysis on the large-scale user interaction log of real recommender system, we model quality effects and user stopping behavior in the recommendation scenario for the first time. Based on the findings, we significantly improve implicit feedback construction. Besides, we study the causes and patterns of user session-level stopping behavior, and propose the model for identifying user stopping intention and even predicting in advance. The proposed model improves the accuracy by 20.3% compared to the baseline model.(2) To address the lack of user complex and dynamic preferences cognitive modeling in existing systems, we model users' preference as a dynamic multi-phase process, and propose an estimation method achieving 36.9% better accuracy compared with the baseline method. We also model the widely existing but implicit negative experience, and propose a model for identifying negative experience with an accuracy of 77.3%. (3) Aiming to bridge the mismatch between current recommender system evaluation and user satisfaction, we propose a new online evaluation framework based on actual preference estimation, which improves the correlation with satisfaction by 20.1% compared to commonly used online metrics. We also propose a new offline evaluation method based on external preference assessment, which significantly alleviates the inconsistency between offline and online evaluation, and leads a new evaluation direction. (4) For the multifaceted evaluation of recommender systems, we model users' perception process of recommendation explanation, and an offline evaluation method for measuring the effectiveness of explanation. We also propose the analysis method for tracing the unfairness origins, and the diagnosis method for detecting unfairness risks.This thesis takes users as the core, deeply models user behavior, cognitive process, and satisfaction, and further solves the important challenges. The work is of great significance in promoting the development of research and application of personalized recommender systems.
