BFRecSys: A Blockchain-based Federated Matrix Factorization for Recommendation Systems

Federated recommendation systems (FRecSys) alleviate the privacy issues of recommendation systems (RecSys) by distributing model training tasks onto users’ local devices. However, they still need a single global server to aggregate training results from users and, thereby are vulnerable to server malfunctioning. Besides, they assume all users voluntarily use their data and computing resources to train recommendation model gradients, which is usually impractical. This paper aims to address the aforementioned problems of FRecSys using blockchain. A blockchain-based federated matrix factorization is designed and realized for RecSys, named BFRecSys. It eliminates the need for a single central server by storing the items and user matrices of the matrix factorization in a blockchain. An incentive mechanism is designed and implemented via smart contracts to record participants’ contributions. Besides, to further enhance the fairness of the incentive mechanism, a fake gradients detection mechanism based on an unsupervised cluster is designed to evict fake gradients in each iteration. The prototype of BFRecSys is realized, and experiments are carried out on public MovieLens datasets and private Ethereum blockchain. The results show that BFRecSys can significantly improve recommendation performance in terms of training accuracy.