Search papers, labs, and topics across Lattice.
This paper addresses the challenge of graph structural entanglement that negatively impacts the performance of graph neural networks (GNNs), particularly at class boundaries where spurious correlations arise from irrelevant neighbors. The authors introduce Boundary Embedding Shaping (BES), an adaptive contrastive learning module that selectively mitigates structural noise at decision boundaries, enhancing classification accuracy without significantly altering model parameters. Experimental results reveal that BES improves GCN performance by an average of 3.3% in node classification and achieves superior accuracy in link prediction tasks compared to existing methods.
Boundary-region entanglement is a critical bottleneck for GNNs, and our adaptive approach boosts classification accuracy by over 3% while maintaining model stability.
Graph neural networks (GNNs) excel at aggregating neighbor information for classification, yet their performance is hindered by graph structural entanglement, where spurious correlations from semantically irrelevant neighbors contaminate node embeddings. This challenge is most acute for nodes near class boundaries in the embedding space, where amplified structural noise blurs decision boundaries and destabilizes predictions. Existing robust GNN methods largely treat all nodes uniformly, ignoring boundary vulnerabilities. In this paper, to improve classification performance, we tackle graph structural disentanglement by identifying boundary-region entanglement as the primary bottleneck and propose Boundary Embedding Shaping (BES), an adaptive contrastive learning GNN plug-in module that selectively suppresses spurious structural noise at decision boundaries with minimal model parameter perturbation. Extensive experiments demonstrate that BES consistently improves boundary discrimination and outperforms existing leading methods. Notably, BES boosts GCN performance by an average of 3.3% in node classification (up to 5.0% on WikiCS) and achieves superior accuracy in link prediction.
