推荐系统已经成为人们在各种场景下信息获取与辅助决策的重要工具。准确理解用户意图是提供精准推荐服务的关键，然而在真实系统中往往只能收集到用户的隐式行为反馈（如点击、浏览等），难以直接获取用户当前的意图。因此，如何充分挖掘隐式反馈背后的用户动态意图是隐式推荐研究的一个重要问题，带来了诸多挑战。首先，用户的真实意图缺少显式反馈，现有推荐模型缺乏对相关影响因素的深入理解与动态建模；其次，隐式反馈数据中缺乏真实负例，基于随机负采样的训练方式难以保证负例选取均符合用户意图；最后，一味以增加隐式反馈交互为目标的推荐结果并不能真正满足用户意图，容易造成标题党等问题。针对以上挑战，本文从三个层面开展了面向用户动态意图的隐式推荐方法研究：在用户建模层面，本文尝试理解用户行为背后的动态意图，挖掘隐式反馈之间的内在关联。首先，本文对同一物品的重复消费行为进行了分析，并通过Hawkes随机过程显式建模其自激励影响。进一步地，考虑到物品间更复杂的相互影响，本文利用知识图谱来帮助理解用户交互之间的相互作用，并通过基于傅里叶变换的频域嵌入实现了不同时间间隔下物品关系动态影响的自适应建模，在亚马逊数据集上较基准算法排序性能提升了13%。在模型学习层面，本文尝试挖掘隐式反馈中的自监督信号，不依赖负采样构建符合用户意图的优化目标。一方面，本文提出基于对齐性与均匀性的用户物品表示学习算法，直接优化推荐中表示的期望性质，显著优于常见随机负采样训练范式。另一方面，本文提出基于意图不变性的序列对比学习算法，能够提升序列表示的质量，作为辅助损失函数结合多种序列推荐模型带来超过10%的性能提升。在系统呈现层面，本文尝试在优化隐式反馈行为的基础上引入物品内容质量对原始推荐列表进行重排序，满足用户对可信推荐的深层次需求。为此，本文提出了内容质量感知的展现调控任务，并设计了基于个性化目标展现质量的重排算法，能够在确保高质量物品得到合理展现的同时，相比基准重排算法更好地平衡推荐系统排序性能、公平性等方面。本文围绕隐式反馈背后的用户动态意图，从用户建模、模型学习、系统呈现三个层面展开研究，所提出的方法涉及个性化推荐反馈循环的各个阶段，能够全面系统地改善真实场景下的推荐性能。

Recommender system has become an important tool for obtaining information and assisting decision-making in various scenarios. To provide precise recommendation services, it is critical to understand user intentions accurately. However, it is common that only implicit feedback data (e.g., click, browse) are available in real-world systems. The current user intention can hardly be obtained directly. Therefore, how to better understand users‘ dynamic intentions behind implicit feedback has become an important issue for implicit recommendation, which brings many challenges. First of all, there is a lack of explicit feedback on users‘ actual intentions, and existing recommendation models lack in-depth understanding and dynamic modeling of relevant influencing factors. Secondly, the implicit feedback data lacks real negative examples, and existing learning algorithms based on random sampling can hardly ensure that all the negative samples are in line with user intentions. Finally, recommendation results that blindly optimize implicit feedback may not cater to users‘ intentions, which can easily lead to problems such as clickbait and domination of low-quality items. To tackle the above challenges, this dissertation conducts research on implicit recommendation methods based on dynamic user intention from three perspectives:From the perspective of user modeling, this dissertation tries to understand the dynamic intentions behind user behaviors, finding internal relationships between implicit feedback. First, this dissertation analyzes the repeat consumption behavior for a single item and models its self-excitation effects through Hawkes process. Further, considering the more complex interplay between different items, this dissertation leverages the knowledge graph to help understand the connections between user interactions. Based on the frequency domain embedding inspired by Fourier transform, this dissertation also achieves the adaptive modeling of the dynamic temporal effects of item relations. The final model improves the ranking performance of the benchmark algorithms by 13% on the Amazon dataset.From the perspective of model training, this dissertation explores self-supervised signals in implicit feedback data, constructing reasonable objectives without relying on negative sampling. On the one hand, this dissertation proposes a representation learning algorithm based on alignment and uniformity, which directly optimizes the desired properties of representations in recommendation. This learning algorithm is significantly better than the common random negative sampling training paradigm. On the other hand, this dissertation proposes a sequential contrastive learning algorithm based on intention invariance modeling to improve the quality of sequence representation. As an auxiliary loss function, the proposed learning algorithm leads to more than 10% performance improvements when combined with various sequential recommendation models.From the perspective of system presenting, this dissertation tries to introduce the item content quality to rerank the original recommendation list based on implicit feedback optimization, aiming to satisfy the user demands for responsible recommendation. In particular, this dissertation proposes the quality-aware exposure regulation task and designs a reranking algorithm based on personalized target exposure quality. Compared with the benchmark reranking algorithms, it can better balance the ranking performance, fairness, and other aspects in recommendation while ensuring the reasonable exposure of high-quality items.Towards dynamic user intention behind implicit feedback, this dissertation conducts research from the perspective of user modeling, model training, and system presenting. The proposed methods involve each stage of the feedback loop in personalized recommendation, which can comprehensively and systematically improve the recommendation performance in real-world applications.