DPMF: Decentralized Probabilistic Matrix Factorization for Privacy-Preserving Recommendation

Collaborative filtering is a popular approach for building an efficient and scalable recommender system. However, it has not unleashed its full potential due to the following problems. (1) Serious privacy concerns: collaborative filtering relies on aggregated user data to make personalized predictions, which means that the centralized server can access and compromise user privacy. (2) Expensive resources required: conventional collaborative filtering techniques require a server with powerful computing capacity and large storage space, so that the server can train and maintain the model. (3) Considering only one form of user feedback: most existing works aim to model user preferences based on explicit feedback (e.g., ratings) or implicit feedback (e.g., purchase history, viewing history) due to their heterogeneous representation; however, these two forms of feedback are abundant in most collaborative filtering applications, can both affect the model, and very few works studied the simultaneous use thereof. To solve the above problems, in this study we focus on implementing decentralized probabilistic matrix factorization for privacy-preserving recommendations. First, we explore the existing collaborative filtering algorithms and propose a probabilistic matrix co-factorization model. By integrating explicit and implicit feedback into a shared probabilistic model, the model can cope with the heterogeneity between these two forms of feedback. Further, we devise a decentralized learning method that allows users to keep their private data on the end devices. A novel decomposing strategy is proposed for users to exchange only non-private information, in which stochastic gradient descent is used for updating the models. Complexity analysis proves that our method is highly efficient with linear computation and communication complexity. Experiments conducted on two real-world datasets FilmTrust and Epinions show that our model gains a guarantee of convergence as the RMSE decreases quickly within 100 rounds of iterations. Compared with the state-of-the-art models, our model achieves lower model loss in rating prediction task and higher precision in item recommendation task