Abstract
Graph Neural Networks (GNN) have attracted increasing attention due to their efficient performance in recommendation systems. However, applying GNNs in session-based recommendations with emerging federated learning (FL) for a privacy-preserving recommendation is challenging. Firstly, constructing a global graph in a centralized manner is forbidden due to the privacy-preserving constraints of FL. Secondly, local graphs in each device contain minimal information on the global graph, causing the inefficient merging of sub-graphs by aggregating local models. Thirdly, the session data in these separated devices are usually extraordinarily non-Independent and Identically Distributed (non-IID), which harms the model performance. In this paper, we bridge the practical gaps between FL and GNN-based session recommendations for the first time by introducing a novel adaptive federated learning method named Federated Graph Aggregation (FedGA). FedGA is beyond the reach of prior adaptive FL methods by incorporating Divergence Resistant Aggregation (DRA) and Conditional Second-Moment Estimation (C-SME), yielding an efficient aggregator where local models trained by the unseen local graph embedding can be efficiently merged. Thanks to the above-proposed strategies, FedGA optimizes models without being interfered with by the aggressive learning rates generated by existing adaptive methods under extreme non-IIDness. In addition, we perform the theoretical analysis of the proposed method, and the results demonstrate that our method achieves a similar rate of convergence as other adaptive FL methods. We validate our method on both open datasets and real-world production data. Results show that our method obtains state-of-the-art performance compared to existing adaptive FL methods while retaining the comparable performance of the centralized methods.