Date of Award

2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Federated Graph Neural Network; Federated Learning; Graph Neural Networks; Saliency Maps

Supervisor

Saeed Samet

Abstract

Graph-structured data present significant challenges in federated learning (FL) environments due to heterogeneous and non-IID data distributions, which often hinder the convergence and effectiveness of Graph Neural Network (GNN) models. This study introduces the Federated Graph Learning Through Saliency Aware Client Clustering (FedSal) framework, designed to enhance FL performance in graph-based contexts by employing saliency maps to identify and emphasize critical data features from local updates. Unlike traditional FL methods that treat all client data uniformly, FedSal facilitates dynamic clustering of clients based on similarities in their saliency data, enabling more tailored and effective model updates. Additionally, we develop FedSal+, which incorporates structural encoding to further enrich saliency maps, refining FedSal's capabilities and enhancing model performance through more nuanced client clustering. Extensive empirical evaluations reveal that FedSal+ achieves accuracy improvements of up to 5.08% over state-of-the-art methods. However, both FedSal and FedSal+ lead to an increase in communication time, indicating a significant trade-off between enhanced accuracy and higher resource consumption. Our findings significantly contribute to advancing federated graph learning by demonstrating that saliency-based FL methods like FedSal and FedSal+ effectively handle data heterogeneity, thereby supporting the development of more personalized and efficient distributed learning frameworks.

Share

COinS