推荐系统与隐私保护研究综述

Survey of recommender system and privacy protection

推荐系统为了能够给用户提供更好的推荐服务,须要收集大量的用户个人信息,在收集这些信息的同时增加了用户隐私泄露的风险.首先,介绍了推荐系统中的关键技术,包括基于协同关系的实体表示学习和基于图模型的实体表示学习;然后,通过对相关研究的归纳和总结,将推荐系统中的隐私保护问题按照用户敏感信息类型进行分类整理,主要分为对用户私有敏感属性的保护、对用户与物品历史交互信息的保护和对用户提交给推荐系统的信息的保护三类;在此基础上,对匿名化、差分隐私、联邦学习和对抗学习四种关键隐私保护技术进行了总结和分析,并重点梳理了这些技术的实现方法、适用场景和优缺点;最后,分析了考虑隐私保护的推荐算法中存在的问题,并尝试给出了未来可能的研究方向.

In order to provide users with better recommendation services,the recommendation system needs to collect a large amount of user personal information,which increases the risk of user privacy disclosure.Firstly,the key technologies in the recommendation system were introduced,including collaborative filtering based entity representation learning and graph based entity representation learning.Secondly,through the induction and summary of relevant research,the privacy protection issues in the recommendation system were classified according to the types of user sensitive information,which are mainly divided into three categories including the protection of users’private sensitive attributes,the protection of historical interaction information between users and items,and the protection of information submitted to the recommendation system by users.On this basis,four key privacy protection technologies,namely anonymization,differential privacy,federated learning and adversarial learning,were summarized and analyzed,and the implementation methods,application scenarios,advantages and disadvantages of these technologies were highlighted.Finally,the problems of privacy preserving recommendation algorithm were analyzed and the possible research directions in the future were given.