BES Module Reduces Boundary Noise in GNN Node Classification

Boundary Embedding Shaping (BES) is a plug-in contrastive learning module for GNNs that targets spurious structural noise specifically at decision boundaries rather than treating all nodes uniformly. It uses adaptive contrastive learning to suppress irrelevant neighbor signals with minimal parameter overhead.