Remedying Coarsening-Based GNN Training under Heterophily via Adaptive Complementary Enhancement

Researchers introduced ACE to make coarsening-based GNN training work better on heterophilic graphs.

The method reconstructs node features, adds anisotropic structural regularization, and balances losses with uncertainty weighting.

ACE helps recover information lost during graph coarsening and improves large-scale GNN training beyond homophilic settings.