S2-HGNN: Scale-Aware Hypergraph Node Classification with Spectral Inductive Bias.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Hypergraph Learning

Background:

• Current hypergraph node classification methods struggle with semi-supervised learning due to reliance on local information and uniform handling of hyperedges.
• Existing approaches fail to effectively integrate global topological context and differentiate modeling strategies for small versus large hyperedges.

Purpose of the Study:

• To propose S2-HGNN, a scale-aware hypergraph neural network framework designed to improve semi-supervised node classification.
• To address limitations in existing methods by incorporating global topological information and employing size-specific hyperedge modeling.

Main Methods:

• Injecting global topological information via complementary hypergraph spectral operators.
• Constructing auxiliary topologies using Top-k constrained clique expansion for small hyperedges and star expansion for large hyperedges.
• Jointly learning node representations from the original hypergraph and auxiliary branches, followed by adaptive fusion for final predictions.

Main Results:

• S2-HGNN consistently outperforms strong baselines on multiple public datasets.
• The proposed method demonstrates superior robustness against feature perturbations.

Conclusions:

• S2-HGNN offers an effective scale-aware approach for semi-supervised hypergraph node classification.
• The framework's ability to leverage global topology and adapt hyperedge modeling leads to improved performance and robustness.