Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs
We investigate the representation power of graph neural networks in the semi-supervised node classification task under heterophily or low homophily, i.e., in networks where connected nodes may have different class labels and dissimilar features. Many popular GNNs fail to generalize to this setting, and are even outperformed by models that ignore the graph structure (e.g., multilayer perceptrons). Motivated by this limitation, we identify a set of key designs -- ego- and neighbor-embedding separation, higher-order neighborhoods, and combination of intermediate representations -- that boost learning from the graph structure under heterophily. We combine them into a graph neural network, H2GCN, which we use as the base method to empirically evaluate the effectiveness of the identified designs. Going beyond the traditional benchmarks with strong homophily, our empirical analysis shows that the identified designs increase the accuracy of GNNs by up to 40% and 27% over models without them on synthetic and real networks with heterophily, respectively, and yield competitive performance under homophily.
PDF Abstract NeurIPS 2020 PDF NeurIPS 2020 AbstractCode
Datasets
Wiki Squirrel
WebKB
Penn94
Cornell
Deezer-Europe
Wisconsin(60%/20%/20% random splits)
Squirrel (60%/20%/20% random splits)
Chameleon (48%/32%/20% fixed splits)
PubMed (60%/20%/20% random splits)
Squirrel (48%/32%/20% fixed splits)
Texas(60%/20%/20% random splits)
Cornell (48%/32%/20% fixed splits)
Cornell (60%/20%/20% random splits)
PubMed (48%/32%/20% fixed splits)
Wisconsin (48%/32%/20% fixed splits)
Cora (48%/32%/20% fixed splits)
Citeseer (48%/32%/20% fixed splits)
Chameleon(60%/20%/20% random splits)
Film(48%/32%/20% fixed splits)
