Personalized federated recommendation system with historical parameter clustering

As an information filtering tool, recommendation system can present interesting contents to specific users through utilizing community users’ information. Due to the increasingly strict collection of user privacy information and improvement of related policies, data are scattered in different organizations as data islands, making it difficult to train a reliable recommendation system model. As for federated learning, an emerging machine learning approach, it enables clients to co-train the model by uploading gradients, which avoids the server to collect sensitive data from clients. To address the problem of not independent and identically distribution in federated learning, we propose a federated recommendation system based on the clustering of historical parameters. The clients perform a weighted average of the historical learning parameters with the global parameters sent by the server through using the time decay factor. The server performs parameter aggregation and clustering on the received parameters. The system performs iterative training based on the users’ historical learning parameters. In addition, when it comes to the problem that the server lacks raw data and cannot provide personalized recommendations for users in the federated recommendation system, we propose a recommendation system model based on user embedding features. The server can use user embedding features for personalized recommendations and it cannot get users’ data through user embedding features. The clients use original data for the local personalized recommendations. We conduct experiments on the real dataset MovieLens-1M. The experimental results show that the proposed federated learning approach is better than the traditional federated learning approach.