Efficient federated graph aggregation for privacy-preserving GNN-based session recommendation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Graph Neural Networks (GNNs) excel in recommendation systems but face challenges in federated learning (FL) for privacy.
• FL constraints prevent global graph construction, and non-IID session data degrades model performance.
• Merging local models from sparse local graphs is inefficient in privacy-preserving scenarios.

Purpose of the Study:

• To introduce a novel adaptive federated learning method, Federated Graph Aggregation (FedGA), for privacy-preserving, session-based recommendations.
• To address the challenges of decentralized graph construction and non-IID data in FL-based GNN recommendations.
• To develop an efficient aggregator for merging local models trained on local graph embeddings.

Main Methods:

• Introduced Federated Graph Aggregation (FedGA), an adaptive FL method incorporating Divergence Resistant Aggregation (DRA) and Conditional Second-Moment Estimation (C-SME).
• Developed an efficient aggregator for merging local models trained on unseen local graph embeddings.
• Incorporated strategies to optimize models without interference from aggressive learning rates under extreme non-IIDness.

Main Results:

• FedGA optimizes models effectively, even under extreme non-IID data conditions.
• Theoretical analysis shows FedGA achieves convergence rates comparable to other adaptive FL methods.
• Empirical validation on open and real-world datasets demonstrates state-of-the-art performance against existing adaptive FL methods.

Conclusions:

• FedGA successfully bridges the gap between FL and GNN-based session recommendations.
• The proposed method achieves state-of-the-art performance while maintaining comparable results to centralized methods.
• FedGA offers an efficient and effective solution for privacy-preserving GNN recommendations in federated settings.