Learning to Collaborate: Personalized Federated GNNs

Mohit Kataria, Nikita Malik, Satyam Govind Shukla, Shreyash Bhilwade, Sandeep Kumar

Keywords: Graph Machine Learning, Federated Learning, Graph Neural Networks, Graph Federated Learning

TL;DR: pFedGNN is a personalized federated graph learning framework that privately infers the global structure and derives a one-shot collaboration graph via coarsening principles, enabling efficient, privacy-preserving, and adaptive client personalization.

Abstract: Federated graph learning enables organizations with distributed graph data to collaboratively train GNNs without sharing raw features. However, in node classification and related tasks, local graphs are typically non-IID in both structure and attributes, making a single global model suboptimal. We propose **pFedGNN**, a personalized federated graph learning framework that *privately infers the global graph structure* across clients and then *derives a client-level collaboration graph* via coarsening principles. Concretely, we estimate a global Laplacian in a privacy-preserving manner and obtain the collaboration graph by lifting with a partition matrix ($L_C = P\,L_G\,P^\top$), where edge weights encode collaboration strength among clients. This structure guides graph-aware parameter aggregation, allowing clients with similar data distributions to share more while preserving personalization. Experiments on diverse graph benchmarks show that pFedGNN significantly improves node classification performance over strong FL/pFL baselines; notably, our method *learns the collaboration graph in one shot*, reducing both communication and computation compared to iterative approaches.

Submission Type: Extended abstract (max 4 main pages).

Submission Number: 72