FedSTG: Breaking through Spatio-Temporal Data Silos with Federated Graph Learning

Spatio-temporal graph (STG) forecasting plays a crucial role in various urban computing applications driven by Web systems. However, conventional STG forecasting methods, which rely on centralized data processing, have posed obstacles to real-world applications in numerous data silo scenarios. Given this, some researchers have preliminarily explored incorporating federated learning into STG forecasting. Though promising, the existing methods encounter two critical issues, i.e., dynamics and heterogeneity in STGs. To this end, we propose a novel personalized <u>Fed</u>erated learning framework for <u>S</u>patio-<u>T</u>emporal <u>G</u>raph forecasting (called FedSTG), along with an evolutionary graph learning module and a personalized federated aggregation algorithm based on evolving temporal patterns in the framework. Our approach enhances the STG learning capabilities in the federated paradigm, yield addressing the problem of STG forecasting in data silo scenarios. Extensive experiments on four real-world datasets in STG forecasting have been conducted to demonstrate the superiority of our approach.