Sub-headline: Researchers from BUPT introduce the RFGDG framework,utilizing RL to dynamically optimize graph generalization in federated settings.

Graph Neural Networks(GNNs)have become pivotal in analyzing complex interactions within social networks and recommendation systems.However,a major technical pain point is the"data silo"problem;high-quality graph data is often strictly controlled by organizations due to privacy and legal requirements,preventing centralized training.While Federated Learning(FL)allows for collaborative training without raw data sharing,it struggles with severe data heterogeneity,where the statistical distributions and graph structures vary significantly across participants.This inconsistency often leads to poor generalization performance when models are deployed in previously unseen domains.
In response,the research team from the School of Computer Science at Beijing University of Posts and Telecommunications developed the Reinforced Federated Graph Domain Generalization(RFGDG)framework.The system introduces a sophisticated feature alignment strategy using random Fourier feature transformation and weighted covariance matrix optimization,which unifies disparate feature representations and reduces redundancy across clients.Locally,it employs GraphSage with an efficient mini-batch graph sampling strategy to maintain structural integrity while minimizing computational overhead.
The core innovation of RFGDG is its dynamic parameter aggregation strategy powered by Deep Reinforcement Learning(specifically the DDPG algorithm).Unlike traditional static weighting,this strategy adaptively adjusts aggregation weights based on the specific contribution of each client and the underlying graph heterogeneity.
Experimental results on multiple datasets,including data from the Weibo platform,show that RFGDG significantly enhances global model accuracy and provides superior adaptability in multi-client environments.This research provides a scalable and robust path for graph-based analysis in privacy-sensitive scenarios.
DOI:10.1007/s11704-025-41295-3