Decentralized neural network based collaborative filtering for privacy concern recommendation systems

Abstract

The growing concern about the privacy of user data is inspiring the development of new privacy preserving machine learning approaches. Decentralized federated learning is such a method which can handle privacy concerns effectively. It consists of servers and clients. Here, a machine learning model is distributed to a number of clients and the clients use their locally stored data to train the model and send the model to the server for aggregation. Here, the client only shares the model parameters with the server. The server receives thousands of locally trained models from clients and performs aggregation to define a global model. Only the model parameters such as weights and biases are being shared between the server and the client. In this paper, we have proposed a decentralized privacy preserving technique for neural collaborative filtering which is widely used in rating prediction for recommendation systems. Here each client receives an initial neural network based collaborative filtering model from the server and trains the model locally with its own data and only sends the model and parameters back to the server for aggregation. This method will eliminate privacy concerns in modern recommendation systems.