User-controlled federated matrix factorization for recommender systems

摘要

Recommendation services have been extensively adopted in various user-centered applications to help users navigate a vast space of possible choices. In such scenarios, data ownership is a crucial concern since users may not be willing to share their sensitive preferences (e.g., visited locations, read books, bought items) with a central server. Unfortunately, data collection is at the basis of modern approaches to the recommendation problem. Decreased users’ willingness to share personal information and data protection policies can result in the “data scarcity” dilemma affecting applications such as recommender systems. In the work at hand, we thoroughly study and extend FPL (Federated Pair-wise Learning), a recommendation approach that follows the Federated Learning principles. In FPL, users collaborate in training a pair-wise learning to rank factorization model while controlling the amount of sensitive data that leaves their devices. An extensive experimental evaluation reveals the effectiveness of the proposed architecture concerning the accuracy and beyond-accuracy objectives and the impact of disclosed users’ information on the quality of the final model. The paper also analyzes the impact of communication costs with the central server on the system’s performance by varying local computation and training parallelism. Furthermore, the study investigates the injection of additional biases in the final recommendation that could affect the fairness of the system. The public implementation is available at https://split.to/sisinflab-fpl.