FedGR: Federated Graph Neural Network for Recommendation Systems

Social recommendation systems based on the graph neural network (GNN) have received a lot of research-related attention recently because they can use social information to improve accuracy and benefits derived from excellent performance in graphic data modeling. A large number studies this area been proposed one after another, but all share common requirement that should be centrally stored. In recent years, there growing concerns about privacy. At same time, introduction numerous stringent protection regulations, represented by general regulations (GDPR), has challenged models with conventional centralized storage. For above reasons, we designed flexible model algorithms for scenarios federated learning. We call it (FedGR). Previous related work only considered GNN, networks, learning separately. Our is first consider three together, carried out detailed design each part. FedGR, used assist modeling implicit vector representation learned users relationship graphs historical item graphs. order protect privacy, FedGR privacy incorporating traditional cryptography encryption techniques “noise injection” strategy, which enables ensure while minimizing loss recommended performance. also demonstrate different paradigm under federation. validated two publicly available popular datasets. According experimental results, decreased MAE RMSE compared previous work, proves its rationality effectiveness.