Hypernetwork-Driven Centralized Contrastive Learning for Federated Graph Classification

Abstract In the realm of Graph Federated Learning (GFL), current methodologies predominantly concentrate on local client data, a focus that often results in a constrained understanding of more expansive, global patterns. This approach typically struggles with Non-IID issues in cross-domain datasets, which impedes the identification of overarching patterns. Contrastive Learning (CL) has emerged as a potent method to improve model’s ability to differentiate between variations across diverse views. However, existing CL frameworks, which are primarily client-centric, fail to fully exploit this advantage. Our empirical findings reveal that the direct application of conventional CL techniques tends to induce homog-enization among clients, a phenomenon particularly pronounced in settings where client datasets exhibit high heterogeneity. To tackle this statistical heterogene-ity and harness inherent data variability, we propose a hypernetwork-based approach, termed CCL. This innovative, server-centric strategy, underpinned by a hypernetwork, adeptly navigates the core challenges associated with traditional client-centric models in the face of heterogeneous datasets. CCL excels in assimilating global patterns derived from multiple clients, effectively capturing a more diverse spectrum of patterns and, consequently, substantially boosting the overall performance of GFL. Our comprehensive experimental evaluations, encompassing supervised, unsupervised, and other harsh scenarios, distinctly affirm CCL’s superiority over prevailing models. Its remarkable compatibility with standard backbones, and the resulting significant enhancements in GFL performance across various contexts, further underscore its effectiveness.