Recommender systems are a fundamental component of contemporary social-media platforms and require feedback submitted from users in order to fulfill their goal. On the other hand, the raise of advocacy about user-controlled data repositories supports the selective submission of data by user through intelligent software agents residing at the user end. These agents are endowed with the task of protecting user privacy by applying a “soft filter” on personal data provided to the system. In this work, we pose the question: “how should the software agent control the user feedback submitted to a recommender system in a way that is most privacy preserving, while the user still enjoys most of the benefits of the recommender system?”. We consider a set of such agents, each of which aims to protect the privacy of its serving user by submitting to the recommender system server a version of her real rating profile. The fact that issued recommendations to a user depend on the collective rating profiles by all agents gives rise to a novel game-theoretic setup that unveils the trade-off between privacy preservation of each user and the quality of recommendation they receive. Privacy is quantified through a distance metric between declared and an “initial” random rating profile; the latter is assumed to provide a “neutral” starting point for the disclosure of the real profile. We allow different users to have different perception of their privacy through a user-dependent utility function of this distance. The quality of recommendations for each user depends on submitted ratings of all users, including the ratings of the user to whom the recommendation is provided. We prove the existence of a Nash equilibrium point (NEP), and we derive conditions for that. We show that user strategies converge to the NEP after an iterative best-response strategy update sequence that involves circulation of aggregate quantities in the system and no revelation of real ratings. We also present various modes of user cooperation in rating declaration, by which users mutually benefit in terms of privacy. We evaluate and compare cooperative and selfish strategies in their performance in terms of privacy preservation and recommendation quality through real movie datasets.

中文翻译：

---

具有自私用户的推荐系统

推荐系统是当代社交媒体平台的基本组成部分，需要用户提交反馈才能实现他们的目标。另一方面，对用户控制的数据存储库的拥护的兴起支持了用户通过驻留在用户端的智能软件代理选择性地提交数据。这些代理具有通过对提供给系统的个人数据应用“软过滤器”来保护用户隐私的任务。在这项工作中，我们提出了一个问题：“软件代理应如何以最能保护隐私的方式控制提交给推荐系统的用户反馈，而用户仍可以享受推荐系统的大多数好处？”。我们考虑了一组这样的代理，每个版本旨在通过向推荐系统服务器提交其真实评分资料的版本来保护其服务用户的隐私。向用户发布推荐取决于所有代理商的集体评分资料，这一事实导致了一种新颖的游戏理论设置，揭示了每个用户的隐私保护与他们收到的推荐质量之间的权衡。隐私通过已声明和“初始”随机评级配置文件之间的距离度量来量化；假定后者为公开真实情况提供了一个“中立”的起点。通过此距离的用户相关效用函数，我们允许不同的用户对其隐私有不同的理解。每个用户的推荐质量取决于所有用户提交的评分，包括向其提供推荐的用户的评分。我们证明了纳什均衡点（NEP）的存在，并为此得出了条件。我们显示，用户策略在迭代的最佳响应策略更新序列之后收敛到NEP，该更新序列涉及系统中总量的循环并且没有实际评级的启示。我们还会在评级声明中介绍用户合作的各种模式，通过这种模式，用户可以在隐私方面互惠互利。我们通过真实电影数据集在隐私保护和推荐质量方面评估和比较合作和自私策略的表现。我们显示，用户策略在迭代的最佳响应策略更新序列之后收敛到NEP，该更新序列涉及系统中总量的循环并且没有实际评级的启示。我们还会在评级声明中介绍用户合作的各种模式，通过这种模式，用户可以在隐私方面互惠互利。我们通过真实电影数据集在隐私保护和推荐质量方面评估和比较合作和自私策略的表现。我们显示，用户策略在迭代的最佳响应策略更新序列之后收敛到NEP，该更新序列涉及系统中总量的循环并且没有实际评级的启示。我们还会在评级声明中介绍用户合作的各种模式，通过这种模式，用户可以在隐私方面互惠互利。我们通过真实电影数据集在隐私保护和推荐质量方面评估和比较合作和自私策略的表现。