Privacy-Preserving Recommendation Based on a Shuffled Federated Graph Neural Network

Benefiting from the efficient extraction of high-order structural information in user–item interaction graphs, graph neural networks (GNNs) are widely used in recommender systems. However, most GNN-based recommenders need to collect user data, which may compromise user privacy. Federated learning is a privacy-preserving framework that can assuage customer privacy concerns by training a shared global model without collecting decentralized private data. To train the GNN-based global recommender on decentralized devices without compromising customer privacy, we propose a novel framework that federalizes the common GNN with a shuffler, which can protect both the ratings and the IDs of items rated by each user. Experiments on real-world datasets illustrate that our framework can achieve more accurate recommendations than state-of-the-art baselines with privacy protection. In particular, the figure improves up to 2.16% and 1.92%, respectively, in terms of the root-mean-square error and mean absolute error.