Can Hypergraph Models Be Strong Baselines For Node-level Tasks? Sloving the Hyperedge Pollution Problem

Shaobo yang, Hong Peng, Xin Jin, Yun Yang, Cheng Xie

Keywords: Graph Neural Networks, Hypergraph Neural Network, Graph Node Classification, Hypergraph Structure Learning

Abstract: Compared to hypergraph-based methods, graph transformers (GTs) and graph neural networks (GNNs)-based models have recently dominated node-level tasks, becoming widely adopted baselines in graph representation learning. However, in this work, HPHNN is proposed to address the Hyperedge Pollution (HP) problem in hypergraph construction, enabling classical hypergraph models to significantly outperform existing GT and GNN models and to serve as strong baselines for node-level graph tasks. Experimental results on 11 real-world graph datasets demonstrate that hyperedge-based models not only outperform these baselines but also exhibit more stable and generalizable performance across diverse node-level tasks by mitigating the HP problem. The findings of this paper challenge the prevailing view that GTs and GNNs possess inherent superiority in node-level graph tasks, establishing another strong baseline model for graph representation learning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 3560