Generating User Privacy-Controllable Synthetic Data for Recommendation Systems

Recommender systems are widely used in e-commerce, news, and advertising, providing personalized recommendations by analyzing user interaction history. However, during large-scale data analysis and sharing, user privacy faces the risk of exposure, especially for users who wish to remain anonymous. While existing synthetic data methods perform well in privacy protection, there is still room for improvement in balancing personalized privacy protection and data utility, particularly in sparse data environments. To address this issue, this paper proposes a user-privacy-controllable data augmentation model that generates synthetic substitutes to replace original data, ensuring anonymity. The model mitigates the data sparsity problem by introducing a global average user-item vector and an attention mechanism. The global vector captures the overall user interest trend, reducing noise and outliers, which allows for a more accurate representation of user interests and a precise selection of items to replace, even in sparse data environments. Additionally, the model dynamically generates synthetic substitutes that align with user interests based on privacy preferences, prioritizing replacing low-interest items, thus achieving a good balance between privacy protection and data utility. Experimental results on three public datasets show that the model effectively protects privacy while maintaining high data utility, outperforming existing mainstream methods in personalized privacy protection.