New Research Challenges GNN Heterophily Classification Explanations

Graph Neural Networks (GNNs) often face difficulties in classifying nodes that exhibit heterophily, meaning they connect to nodes of different classes. Traditional explanations for this failure have pointed to either the rarity of such nodes within the dataset or inherent limitations in how GNNs aggregate neighborhood information. This new study critically examines these two perspectives. The researchers evaluated six different GNN architectures across five datasets with varying degrees of homophily. Their findings suggest that homophilic nodes are generally easier to classify, even when they are rare, which contradicts the idea that rarity is the main problem. Furthermore, the study indicates that the necessary information for correctly classifying heterophilic nodes is often present within the GNN's learned representations and can be recovered by simply retraining the final classification layers.