Privacy-preserving recommender system with VAEs : optimization, privacy, and application

Abstract

Recommender Systems (RS) have been widely applied in multiple industrial services, which gained a lot of attention from both industry and academia as a core component of item recommendation for users. A typical RS collects personal data, analyzes the relations among users and items, and recommends potential item candidates within users' interests. To optimize the RS, variational autoencoders (VAEs) have been adopted to provide more precise recommendations. VAE is a non-linear model powered by neural networks and can capture more complex input data patterns. However, personal information, such as user profile, user preference, and feedback, is sensitive to each user if adversaries maliciously used the data for further illegal analyses. To provide better privacy protection for user information, differential privacy (DP) has been widely applied in RS, ensuring privacy protection over a group of participants by adding noise into the training process. Recent studies discovered that the trained model is controlled by large user groups, which is known as the ``disparate effect''. And some researchers even observed that DP could even worsen the disparate effect of performance degradation, while most prior studies applied the same privacy guarantee for each user and neglected the disparate effect on the model performance. To solve those issues, we first propose user-level differential privacy, which each user chooses within its interest. Then we analyze the relation of RS model performance and dataset size and suggest an efficient way of building the dataset for training RS. Extensive experimental results on diverse datasets verify our solutions' effectiveness, where we compare our proposed model with multiple standard models and state-of-the-art DP-RS designs. Furthermore, we apply the RS into an exercise video recommendation app for Parkinson's disease (PD) patients, where we implement the recommendation algorithm on the back-end server that provides high efficiency and achieves good recommendation accuracy.