Efficient and Reliable Federated Recommendation System in Temporal Scenarios

Abstract

Addressing privacy concerns and the evolving nature of user preferences, it is crucial to explore collaborative training methods for federated recommendation models that match the performance of centralized models while preserving user privacy. Existing federated recommendation models primarily rely on static relational data, overlooking the temporal patterns that dynamically evolve over time. In domains like travel recommendations, factors such as the availability of attractions, introduction of new activities, and media coverage constantly change, influencing user preferences. To tackle these challenges, we propose a novel approach called FedNTF. It leverages an LSTM encoder to capture multidimensional temporal interactions within relational data. By incorporating tensor factorization and multilayer perceptrons, we project users and items into a latent space with time encoding, enabling the learning of nonlinear relationships among diverse latent factors. This approach not only addresses the privacy concerns by preserving the confidentiality of user data but also enables the modeling of temporal dynamics to enhance the accuracy and relevance of recommendations over time.