Revisiting Classic Graph Neural Networks for Hypergraph Node Classification: An Extensive Empirical Study

Abstract

Have we been too quick to dismiss classic Graph Neural Networks (GNNs) as inadequate for hypergraph learning? Despite the rise of Hypergraph Neural Networks (HNNs) as the dominant framework for modeling high-order relations, classic GNNs have often been regarded as inherently inferior. However, their underperformance may have resulted from evaluations using overly simplified configurations that omit essential training techniques. To reassess their potential, we conduct an extensive empirical comparison for node classification on 14 real-world hypergraphs, evaluating classic GNNs with clique and star expansions alongside standard training techniques, including dropout, normalization, residual connections, and DropEdge. When properly tuned, classic GNN baselines achieve mean accuracy comparable to or higher than the best HNN baseline on 12 of the 14 datasets, spanning both homophilous and heterophilous settings. Our ablation studies further show that star expansion is particularly effective on heterophilous hypergraphs, the optimal network depth depends on the level of homophily, and dropout provides the largest individual performance gains. These findings demonstrate that classic GNNs remain strong baselines for hypergraph learning and provide practical guidance for effectively configuring them across diverse hypergraph settings.