Robust Graph Neural Networks Under Noisy Graphs: A Comparative Study of Adaptive and Learnable Edge Reliability

This study compares how well different types of GNNs hold up when the underlying graph is noisy. Standard architectures like GCN, GAT, and GraphSAGE are tested alongside three edge-reliability mechanisms: one based on feature similarity, one that adds structural information, and one that learns the weighting directly. All are evaluated under the same controlled noise conditions across Cora, CiteSeer, and PubMed. Most existing work in this area tests these kinds of mechanisms separately, often on different datasets or under different noise setups, so it’s hard to know how they’d actually compare side by side. Here, all six models see identical corruption at each noise level, which makes a fair comparison possible. The results show that edge-reliability mechanisms generally help, but not evenly. No single method comes out on top across all three datasets. The learnable mechanism, for instance, performs strongly on PubMed, retaining nearly 99% of its clean accuracy at 30% noise, but it does noticeably worse on Cora. This suggests that choosing a robustness mechanism should depend on the characteristics of the graph itself, rather than assuming one approach works best everywhere.