Federated graph learning with spatio-temporal dynamics for cross-border recommendation

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Cross-border data sharing for recommendation systems is hampered by privacy regulations and data scarcity.
• Existing federated graph neural network methods struggle with heterogeneous data due to user isolation and business homogeneity.
• Traditional federated averaging (FedAvg) is insufficient for complex, real-world recommendation scenarios.

Purpose of the Study:

• To propose FedSTAR, a novel privacy-preserving framework for cross-border recommendation systems.
• To address the limitations of existing federated learning methods in handling heterogeneous data.
• To enhance recommendation accuracy and robustness in cross-border data sharing scenarios.

Main Methods:

• Integration of spatio-temporal dynamic modeling with federated graph neural networks.
• Development of a dynamic sequential graph structure to model evolving user preferences.
• Implementation of a multi-head attention mechanism for spatial neighbor filtering.
• Introduction of a personalized federated aggregation strategy to replace FedAvg for adaptive data fusion.

Main Results:

• FedSTAR achieved average improvements of 2-5 percentage points in Recall@20.
• FedSTAR demonstrated average improvements of 1-3 percentage points in NDCG@20.
• The framework proved effective on Gowalla, Yelp 2018, and Amazon Book datasets.

Conclusions:

• FedSTAR offers a secure and practical solution for cross-border recommendation systems.
• The framework provides high accuracy and robustness under privacy constraints (model updates only).
• Personalized federated aggregation effectively fuses heterogeneous multi-source data.