TAGR: Lightweight graph repair boosts GNN robustness against noise

Graph neural networks (GNNs) rely on high-quality graph topology, but real-world graphs often contain noisy or missing edges. Existing approaches either delete suspicious edges (which cannot recover missing connections) or learn a new graph structure (which adds complexity). In their arXiv paper, Anubha Goel and Juho Kanniainen introduce Topology-Aware Gaussian Repair (TAGR), a lightweight framework that repairs the graph before message passing. TAGR constructs a sparse feature-neighborhood graph using an adaptive Gaussian kernel to add auxiliary edges between similar nodes, then applies a topology-aware residual correction that reweights the original edges based on local feature and structural consistency. The result is a repaired graph that can be used with any standard GNN without modifying its architecture.

Extensive experiments on citation networks (e.g., Cora, Citeseer, Pubmed) show that TAGR significantly improves GNN robustness when up to 50% of edges are noisy or when 30% of edges are missing. The Gaussian feature-neighborhood repair provides the primary robustness gain, while the residual correction stabilizes performance on incomplete graphs. Notably, TAGR achieves these results with a sparse graph, avoiding the computational overhead of dense adjacency matrix learning. This makes TAGR a practical drop-in solution for deploying GNNs in applications where graph data is inherently imperfect.