互联网信息的爆发式增长给用户带来了严重的信息过载。个性化推荐系统可以根据用户的喜好与需求筛选信息，提高用户的信息获取效率。现有的个性化推荐系统研究大多着眼于提升推荐的准确性与用户黏性等商业性指标。但是，由于推荐系统存在一定的社会效应，其背后的价值观很大程度上决定了其长期的社会影响。因此，负责任的个性化推荐系统是信息检索领域的重要研究方向。构建负责任的个性化推荐系统面临以下挑战：（1）如何从去中心化的用户数据高效准确地训练隐私保护的推荐模型；（2）如何帮助数据海量而嘈杂的推荐系统防御内容安全威胁；（3）如何避免推荐模型拟合用户的敏感属性信息，减轻推荐的不公平性；（4）如何消除推荐环路中的隐含偏见以免被算法放大。本文围绕这些问题，对推荐系统的隐私安全、内容安全、推荐公平和偏见消除等方面展开研究，为推荐算法治理提供了理论方案与关键技术。本文的研究内容和创新点如下： 第一，提出了一种高效通信的隐私保护推荐模型的学习方法。该方法解耦联邦学习中的知识抽取与知识迁移机制，采用自适应互知识蒸馏的机制实现知识的高效传递，并利用动态梯度分解的机制进一步降低通信开销。该方法在推荐模型准确性接近无损的前提下，降低去中心化推荐模型训练中94.6%的通信开销。 第二，提出了一种推荐系统的观点失真攻击与防御方法。该攻击方法揭示了内容推荐系统可能面临的一种新型安全风险，验证了攻击者操控推荐结果的观点分布与干预用户信息认知的可能。在此基础上，本文提出了一种针对性的防御方法来加固各类内容推荐算法，将该类恶意攻击带来的影响平均降低81.2%。 第三，提出了一种基于分解对抗学习的公平推荐方法。该方法将用户模型分解为接近正交的两个部分，分别用于建模用户的敏感属性信息与兴趣信息。其中一部分通过属性预测来增强敏感属性建模，另一部分利用对抗学习移除敏感属性信息。该方法实现了仅用0.94%的绝对性能损失换取与随机推荐接近的公平性。 最后，发现了推荐模型隐藏的情感偏差放大回路现象，并提出了一种模型训练方法进行消除。本研究发现推荐模型会继承和放大数据中隐藏的用户情感偏见，从而自发操控推荐结果的情感倾向。为消除这一现象，本研究提出将情感信息与情感无关的内容偏好信息解耦，根据情感无关的信息计算纠偏的排序分数。该方法能够以较小的性能损失为代价，减少推荐结果中97.3%的情感偏差。

The explosion of online information brings heavy information overload to users. Personalized recommender systems can filter information based on users‘ preferences and needs to improve their information acquisition efficiency. Existing studies on personalized recommender systems usually aim to promote business metrics such as recommendation accuracy and user engagement. However, due to the societal effects of recommender systems, their long-term impacts on society heavily depend on their values behind. Therefore, responsible and personalized recommender system is an important research direction in the information retrieval field. The construction of responsible personalized recommender systems faces the following challenges: (1) how to efficiently learn accurate and privacy-preserving recommendation models from decentralized user data; (2) how to secure recommender systems with big but noisy data to defend against content security threats; (3) how to prevent the models from fitting users‘ sensitive attributes to mitigate recommendation unfairness; and (4) how to avoid the algorithmic amplification of underlying biases in the loop of recommendation. We study several aspects of recommender systems to address these challenges, including privacy security, content security, recommendation fairness, and debiasing, to offer theoretical blueprints and key techniques for the governance of recommendation algorithms. The contributions of this dissertation are as follows: First, we propose a communication-efficient method for privacy-preserving recommendation model training, which decomposes the knowledge extraction and transfer mechanisms in federated learning. In this method, we propose an adaptive mutual knowledge distillation method to achieve effective knowledge exchange, and use a dynamic gradient factorization mechanism to further reduce the communication cost. The proposed method reduces 94.6% of communication cost in decentralized recommendation modeling training without substantial performance loss. Second, we propose an opinion distortion attack and defense method in recommender systems. The attack method reveals a new type of vulnerability in content recommender systems, showing the possibility of manipulating the opinion distribution in the recommendation results and interfering user perception. Accordingly, we propose a defense method to consolidate various content recommendation algorithms, which mitigates 81.2% of the malicious attack impact in average. Third, we propose a fair recommendation method based on decomposed adversarial learning. It decomposes the user model into two almost orthogonal parts to model sensitive user attribute information and interest information, respectively. The former part uses an attribute prediction task to strengthen its ability of sensitive attribute modeling, and the latter part uses adversarial learning to remove the information of sensitive attributes. This method can achieve similar recommendation fairness as random ranking by only sacrificing 0.94% of recommendation performance. Finally, we discover the phenomenon of sentiment bias amplification loops hidden in recommender systems and propose a model training method to remove them. We find recommendation models may inherit and magnify users‘ sentiment biases encoded in data to spontaneously manipulate the sentiment orientation of recommendation results. To eliminate this phenomenon, we propose to disentangle sentiment-independent content and preference information from sentiment information, and generate debiased ranking scores based on sentiment-independent information. This method can remove 97.3% of the sentiment biases in the recommendation results with a minor performance loss.