PEPPER

Recommender systems are proving to be an invaluable tool for extracting user-relevant content helping users in their daily activities (e.g., finding relevant places visit, consume, items purchase). However, effective, these need collect and analyze large volumes of personal data location check-ins, movie ratings, click rates .. etc.), which exposes numerous privacy threats. In this context, recommender based on Federated Learning (FL) appear a promising solution enforcing as they compute accurate recommendations while keeping the users' devices. FL, therefore FL-based systems, rely central server that can experience scalability issues besides being vulnerable attacks. To remedy this, we propose PEPPER, decentralized system gossip learning principles. model updates aggregate them asynchronously. At heart PEPPER reside two key components: personalized peer-sampling protocol keeps neighborhood each node, proportion nodes have similar interests former simple yet effective aggregation function builds is better suited user. Through experiments three real datasets implementing use cases: check-in recommendation recommendation, demonstrate our converges up 42% faster than with other solutions providing 9% improvement average performance metric such hit ratio 21% long tail compared competitors.