Federated Graph Neural Networks for Heterogeneous Graphs with Data Privacy and Structural Consistency

This paper addresses the problem of joint modeling for multi-source heterogeneous graph data in distributed environments by proposing a federated graph neural network classification framework driven by structural alignment and consistency regularization. The method preserves data locality by enabling each participant to learn node features and topological information through a local graph neural network encoder. A cross-source structural alignment module maps embeddings from different graphs into a shared representation space, mitigating semantic inconsistencies caused by structural differences. Additionally, a consistency regularization mechanism is introduced to enhance the robustness of node representations through multi-view perturbations, improving the model's generalization ability during training. At the global level, a federated averaging strategy is adopted to periodically aggregate local models, enabling collaborative optimization and enhancing the consistency and discriminative capacity of the global representation. To validate the effectiveness of the proposed approach, experiments are conducted in a multi-source heterogeneous graph environment using various node distribution strategies. The results show that the method outperforms existing federated graph learning models in terms of accuracy, clustering consistency, and structural expressiveness, achieving efficient multi-source graph classification while preserving data privacy.