FedPPD: Towards effective subgraph federated learning via pseudo prototype distillation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Subgraph federated learning (subgraph-FL) enables collaborative training of graph neural networks (GNNs) across clients.
• Real-world subgraph-FL is challenged by subgraph heterogeneity, leading to performance degradation of the global model.
• Existing methods often rely on parameter aggregation, oversimplifying subgraph knowledge and resulting in suboptimal inference.

Purpose of the Study:

• To propose a novel method, FedPPD (Federated Pseudo Prototype Distillation), to address subgraph heterogeneity in subgraph-FL.
• To improve the performance of global GNN models in decentralized graph learning scenarios.
• To effectively transfer rich subgraph knowledge to a global model without oversimplification.

Main Methods:

• FedPPD employs a generator guided by local prototypes to explore the global input space.
• Pseudo graphs are generated to distill knowledge from local GNNs to the aggregated global GNN.
• This process aims to convey nuanced subgraph information lost during standard aggregation.

Main Results:

• FedPPD consistently outperforms state-of-the-art baselines in subgraph-FL tasks.
• Experimental validation across six public datasets confirms the effectiveness of the proposed method.
• The pseudo prototype distillation approach enhances the global model's ability to leverage diverse subgraph data.

Conclusions:

• FedPPD effectively mitigates the negative impact of subgraph heterogeneity in federated learning.
• The proposed knowledge distillation strategy offers a superior alternative to simple parameter aggregation for global GNN inference.
• FedPPD represents a significant advancement in achieving robust and high-performing subgraph-FL systems.