Forget Me If You Can: Auditing User Data Revocation in Recommendation Systems

• 成果类型: Article; Early Access
• 署名作者: Zhu, Zhihao; Yang, Yi; Fan, Yangyang; Lian, Defu
• 署名单位: Hong Kong University of Science & Technology; Hong Kong Polytechnic University; Chinese Academy of Sciences; University of Science & Technology of China, CAS
• 刊物名称: INFORMATION SYSTEMS RESEARCH
• ISSN/ISSBN: 1047-7047; 1526-5536
• DOI: 10.1287/isre.2024.1179
• 发表日期: 2026-03-30
• 关键词: the right to be forgotten; data revocation auditing; profiling; privacy-preserving; sequential recommendation systems; membership inference; information privacy research; records; state

摘要

Recommendation systems have become integral to modern e-commerce and streaming platforms, enhancing user experience through personalized content and product suggestions. Whereas users benefit from personalized recommendations by allowing platforms to leverage their interaction data for model training, regulations such as the General Data Protection Regulation (GDPR) uphold users' right to be forgotten, which permits individuals to request the deletion of their personal data, including interaction histories. Although removing user data from storage systems is relatively straightforward, it is equally critical to eliminate user-specific behavioral patterns from the trained recommendation model. Otherwise, the model may continue to produce recommendations that reflect a user's past behaviors or preferences, resulting in continued profiling even after a data deletion request. In this work, we address the novel design problem of user data revocation auditing and propose a method, named RecAudit, to examine whether a sequential recommendation system has effectively forgotten, or continues to retain, an individual user's behavioral data after a deletion request. Extensive experiments on multiple realworld data sets demonstrate that RecAudit substantially outperforms existing auditing and membership inference baselines across a wide range of settings. We also examine auditing performance when machine unlearning, an increasingly practical approach for removing data from trained models, is applied. As an auditing tool, RecAudit can help identify high-risk users whose data have not been properly forgotten, thereby facilitating targeted model unlearning and enhancing overall user privacy preservation. Our study contributes to information systems research by developing a privacy-preserving IT artifact that operationalizes regulatory requirements on data revocation and profiling in learningbased recommender systems.