Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs.

摘要: We investigate the representation power of graph neural networks in semi-supervised node classification task under heterophily or low homophily, i.e., where connected nodes may have different class labels and dissimilar features. Many popular GNNs fail to generalize this setting, are even outperformed by models that ignore structure (e.g., multilayer perceptrons). Motivated limitation, we identify a set key designs -- ego- neighbor-embedding separation, higher-order neighborhoods, combination intermediate representations boost learning from heterophily. combine them into network, H2GCN, which use as base method empirically evaluate effectiveness identified designs. Going beyond traditional benchmarks with strong our empirical analysis shows increase accuracy up 40% 27% over without on synthetic real heterophily, respectively, yield competitive performance homophily.