New Framework Combats Oversmoothing in Hypergraph Neural Networks

Hypergraph Neural Networks (HGNNs) offer enhanced expressive power through higher-order couplings, but this also exacerbates the problem of representation collapse, or oversmoothing, in deeper networks due to intense multi-way feature mixing. This study analyzes hypergraph oversmoothing from a dynamical-systems perspective, viewing message passing as an incidence-level diffusion process. The analysis reveals that pure diffusion exponentially contracts node representations and drives the Dirichlet energy to zero, indicating oversmoothing as an intrinsic energy dissipation phenomenon. To counteract this, the researchers propose Hypergraph Neural Reaction-Diffusion (HNRD), which integrates a reaction mechanism. This mechanism acts on the transverse component of representations to compensate for diffusion-induced dissipation, thereby stabilizing discriminative features. The framework ensures global well-posedness and proves that the null-mode-free Dirichlet energy remains bounded. A practical HNRD layer is derived via forward-Euler discretization, demonstrating stable performance in deep propagation. Experiments show HNRD consistently outperforms existing hypergraph baselines, maintaining stable performance and non-zero Dirichlet energy even with increased depth and perturbations.