基于Rényi差分隐私的图卷积协同过滤推荐算法  

Graph Convolutional Collaborative Filtering Recommendation Algorithm Based on Rényi Differential Privacy

作　　者：王锟 王永[1,2] 刘金源 邓江洲 WANG Kun;WANG Yong;LIU Jin-Yuan;DENG Jiang-Zhou(College of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Economics and Management,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Education Information Technology Center,Southwest University of Political Science and Law,Chongqing 401120,China)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]重庆邮电大学经济管理学院,重庆400065 [3]西南政法大学教育信息技术中心,重庆401120

出　　处：《软件学报》2025年第3期1202-1217,共16页Journal of Software

基　　金：国家自然科学基金(62272077);重庆市自然科学基金(cstc2021jcyj-msxmX0557);教育部人文社科规划项目(20YJAZH102)。

摘　　要：近年来,图卷积网络作为一种强大的图嵌入技术在推荐系统领域得到广泛应用.主要原因是推荐系统中大多数信息可以建模为图结构,而图卷积网络是一种基于图结构的深度学习模型,有助于挖掘图数据中用户和项目之间的潜在交互,从而提高推荐系统的性能.由于推荐系统的建模通常需要收集和处理大量的敏感数据,因此可能会面临隐私泄露的风险.差分隐私是一种具有坚实理论基础的隐私保护模型,已被广泛应用于推荐系统中解决用户隐私泄露的问题.目前基于差分隐私的研究主要是面向独立同分布的数据模型.然而,在基于图卷积网络的推荐系统中,数据之间关联性强且不具有独立性,这使得现有方法难以对其进行有效的隐私保护处理.为解决该问题,提出基于Rényi差分隐私的图卷积协同过滤推荐算法RDP-GCF,旨在保护用户与项目交互数据安全的前提下,实现隐私性和效用性之间的平衡.该算法首先利用图卷积网络学习用户/项目的嵌入向量;然后,采用高斯机制对嵌入向量进行随机化处理,同时基于采样的方法放大隐私预算,减少差分噪声注入量,以提升推荐系统的性能;最后,通过加权融合的方式得到用户/项目的最终嵌入向量,并应用于推荐任务.在3组公开数据集上进行实验验证.结果表明,与现有同类方法相比,所提算法能更好地实现隐私保护与数据效用之间的平衡.Recently,graph convolutional network(GCN),as a powerful graph embedding technology,has been widely applied in the field of recommendation.The main reason is that most of the information in recommender systems can be modeled as graph-structured data,and GCN,as a deep learning model that operates on graph structures,helps to explore the potential interactions between users and items in graph-structured data,to enhance the performance of the recommender systems.Since the modeling of recommender systems usually needs to collect and process a large amount of sensitive data,it may face the risk of privacy leakage.Differential privacy,as a privacy protection model with a solid theoretical foundation,has been widely used in recommender systems to solve the problem of personal privacy leakage.Currently,the research based on differential privacy is mainly oriented to independent and identically distributed data models.However,data within GCN-based recommender systems is highly correlated and not independent,making the existing privacy protection methods less effective.To solve the problem,this study proposes a graph convolutional collaborative filtering recommendation algorithm based on Rényi differential privacy(RDP-GCF for short),aiming to achieve a balance between privacy protection and utility while ensuring the security ofuser-item interaction data.The algorithm first utilizes GCN techniques to learn the embedding vectors for users and items.Then,the Gaussian mechanism is used to randomize the embedding vectors,and a sampling-based method is used to amplify the privacy budget and minimize the injection of differential noise,thereby improving the performance of the recommender system.Lastly,the final embedding vectors of the users and items are obtained by a weighted fusion and applied to the recommendation tasks.The proposed algorithm is validated through experiments on three publicly available datasets.The results show that compared to existing similar methods,the proposed algorithm more effectively achieves a balan

关 键 词：推荐系统 协同过滤 图卷积网络 隐私保护 Rényi差分隐私 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]