On the Geometric Coherence of Global Aggregation in Federated GNN
New server-side method prevents destructive interference when merging AI models trained on different graph structures.
Researchers Chethana Prasad Kabgere and Shylaja SS propose GGRS (Global Geometric Reference Structure), a server-side framework for Federated Graph Neural Networks (GNNs). It addresses a 'geometric failure mode' where standard aggregation of client updates degrades the model's relational reasoning, a problem not shown in standard loss metrics. GGRS regulates updates using geometric criteria to preserve message-passing coherence without accessing private client data, demonstrated on datasets like Amazon Co-purchase.
Why It Matters
Enables more reliable, privacy-preserving AI training on distributed, interconnected data like social networks or supply chains.