Enhanced federated contrastive learning of graph-level representations with neighborhood consistency and personalized aggregation

Abstract

This dissertation investigates the challenges of learning high-quality graph-level representations in federated settings while preserving data privacy. Graph neural networks have emerged as powerful tools for analyzing graph-structured data, while federated learning enables collaborative model training without sharing raw data. However, the combination of these approaches faces challenges including non-IID data distributions and structural information loss during aggregation. Building upon the Federated Contrastive Learning of Graph-Level Representations (FCLG) framework, this research proposes two enhancements: Neighborhood Consistency Aggregation (NC) and Personalized Parameter Aggregation (PA). NC preserves local topological structures by introducing a regularization term that encourages connected nodes to maintain similar representations. PA addresses non-IID challenges through a two-stage aggregation process that balances new client updates with historical knowledge. Experiments on four benchmark datasets (PROTEINS, ENZYMES, DHFR, and NCI1) demonstrate improved performance over baseline methods. The NC approach shows consistent gains across all datasets, particularly on molecular graphs, while PA demonstrates increased benefits in scenarios with higher data heterogeneity. This research contributes to the advancement of privacy-preserving graph learning by addressing both client-side structure preservation and server-side aggregation challenges. The proposed methods enable more effective collaboration on graph-based machine learning tasks across distributed data sources, particularly in settings where data cannot be centralized due to privacy constraints or computational limitations.