Securing recommendation systems in federated learning based on neural collaborative filtering method with a pairwise approach

Recommendation systems are indispensable for delivering personalized recommendations to users, yet the escalating concerns regarding data privacy have necessitated the exploration of secure solutions. This work introduces an innovative algorithm designed for training a Collaborative Filtering recommendation system within a Federated Learning framework, specifically optimized for ranking tasks. By leveraging cutting-edge matrix factorization techniques such as Neural Collaborative Filtering and Bayesian Personalized Ranking, our approach aims to preserve user interaction information while ensuring privacy protection. Unlike traditional methods that entail transferring all interaction data, our algorithm minimizes communication costs by only requiring information on a subset of interactions per training iteration. Our extensive experiments demonstrate the remarkable efficacy of the newly proposed model in harnessing user data, surpassing the state-of-the-art FedeRank. Our model achieves not only a commendable hit ratio of 62% but also the highest quality of recommended items, boasting an impressive Normalized Discounted Cumulative Gain score. This breakthrough represents a significant stride towards harmonizing performance and privacy in Recommendation Systems, pivotal for instilling user confidence and satisfaction in the modern digital era.