Reproduce “Improving Graph Neural Networks with Simple Architecture Design” on OpenResearch

About this paper

Sunil Kumar Maurya, Xin Liu, Tsuyoshi Murata

Researchers from Tokyo Institute of Technology and AIST developed Feature Selection Graph Neural Network (FSGNN), a two-layer GNN that decouples feature generation from representation learning and employs soft-selection and hop-normalization. The model achieved up to 64% higher accuracy on heterophilous datasets like Squirrel, maintained competitive performance on homophilous graphs, and scaled effectively to large datasets such as ogbn-papers100M with a 67.17% accuracy.

Key insights

• Decoupling feature generation enables GNNs to effectively handle both homophilous and heterophilous graphs by selectively weighting features from different hop distances.
• The soft-selection mechanism provides interpretability, showing that self-loop features are crucial for homophilous graphs, while ego and first-hop features without self-loops are important for certain heterophilous datasets.
• Hop-normalization improves discriminative power, leading to more separated node clusters in embedding space and better classification accuracy, especially on heterophilous graphs.

Summary by alphaXiv.