Learning Hierarchy-Aware Federated Graph Embedding for Link Prediction

Abstract

As a key task in graph data mining, graph link prediction plays a vital role across various downstream applications, including e-commerce recommendations, protein interaction fore-casts, financial fraud detection, and discovering potential social relationships. Traditionally, methods for graph link prediction rely on a single data source to learn graph embeddings. However, in situations where data isolation is exacerbated by privacy regulations like GDPR, performance can decline due to relying on limited local data. To address this challenge, we propose Federated Hierarchy-aware Graph Embedding (FedHGE), which facilitates hierarchical graph embedding while ensuring data remains securely within its local repository. FedHGE incorporates two main schemes: graph hierarchical embedding and federated link prediction. Graph hierarchical embedding leverages neural networks to extract pattern representations from the graph at different levels of granularity, resulting in more comprehensive embeddings. Federated link prediction allows model training across decentralized data sources so that each party only shares optimized model parameters instead of raw data. Experimental results demonstrate that FedHGE outperforms state-of-the-art methods for graph link prediction on several benchmark datasets.