User-aware differential privacy recommendation framework based on dual variational autoencoders

With the rapid development of information technology and the continuous expansion of data scale, it’s becoming increasingly challenging to efficiently obtain interesting content from massive information sources. Recommender systems (RS) typically rely on the collection and analysis of personal profiles, preference records, and interaction feedback, which may lead to severe privacy leakage risks due to improper operations or malicious exploitation. To address these concerns, the differential privacy (DP) mechanism has been increasingly incorporated into RS by imposing privacy-preserving constraints, thereby protecting sensitive information from leakage. In practice, the differential privacy mechanism is promising yet underexplored to achieve the balance in recommender performance and privacy protection. Here, we propose a novel user-aware differentially private recommender system based on Dual Variational Autoencoders (VAE), referred to as DP-DVAE. Specifically, DP-DVAE constructs a dual-VAE architecture to jointly learn latent representations for users and items, which will be further integrated with matrix factorization. This strategy enables effective dimensionality reduction while enhancing the model’s capacity to capture intricate latent feature distributions. During the training phase, the differential privacy mechanism is implemented by injecting extra noise, allowing DP-DVAE to maintain high performance in recommendation while ensuring privacy preservation. Moreover, we leverage the metadata to construct user-level prior distributions, which can extract semantic information and significantly enhance the performance of the Variational Autoencoders. Extensive experimental results on multiple datasets demonstrate that DP-DVAE is capable of delivering high-quality recommendations in contrast to the benchmark methods, as well as ensuring high-performance privacy preservation.