Fission-based Dynamic Hypergraph Neural Network

doi:10.1016/j.patcog.2026.113995

Fission-based Dynamic Hypergraph Neural Network

2026-05-01

0

PRE

AI

DOI：10.1016/j.patcog.2026.113995

Original

Original request for help

Save

Abstract

En 中文

The performance of Hypergraph Neural Networks (HGNNs) largely depends on the quality of the hypergraph structure. However, generating high-quality structures remains challenging, and their homogeneity assumptions affect their effectiveness. Dynamic Hypergraph Neural Networks (DHGNNs) aim to capture more hidden information by optimizing the hypergraph structure. However, most existing DHGNNs cannot adjust the number of hyperedges in the hypergraph and the number of hyperedge connections of the vertices, limiting their ability to fully exploit the hypergraph structure. This paper presents a novel method called the Fission-based Dynamic Hypergraph Neural Network (FDHNN). FDHNN enhances performance by dynamically adjusting the hypergraph structure and optimizing hypergraph convolution. Specifically, the method first performs controlled fission on the graph structure to reveal the multiple views information of the vertices. Then, vertices are embedded into different hyperedges through collaborative networks and fission graphs, thereby constructing a hypergraph that is softly associated with refined features and structures. We design an effective hypergraph convolution algorithm that uses wavelets to extract high-frequency and low-frequency information from the data. Additionally, we explore the impact of topological structure on HGNNS in low homophily data. Experiments on real datasets validate the effectiveness of the algorithm.

Keywords:

Dynamic hypergraph

Hypergraph neural network

Hypergraph convolution

Node fission

Wavelet transform

Fission-based Dynamic Hypergraph Neural Network

Abstract

Journal

Organization

Cited Papers