A Simple and Efficient Federated Recommender System

Abstract

Federated Learning (FL) is recently explored as a machine learning paradigm to communally gain generalizable knowledge from the data available in a collection of edge devices without the requirement to transfer the data. FL gives rise to the opportunity to train models on edge devices while preserving user's privacy as the data never leaves user's premises. In this paper, we introduce a simple yet efficient extension of FL for recommender systems to improve on personalization and discuss closely-related meta-learning algorithms. Compared to state-of-the-art federated recommenders, our proposed algorithm is simpler and more robust in real-life scenarios. Through experiments on benchmark data, we evaluate our algorithm in root mean squared error (RMSE) of user's rating prediction.