Heterophily-aware dynamic hypergraph for semi-supervised classification

Hypergraph neural networks, as high-order graph neural networks, excel in handling intricate relationships within non-Euclidean infinite-dimensional spaces. However, conventional homophily assumption-based hypergraph methods exhibit limited effectiveness in semi-supervised classification scenarios involving heterophily problem, where neighboring nodes often belong to dissimilar categories. To address this challenge, this paper proposes a Heterophily-Aware Dynamic Hypergraph (HADHG) framework grounded in heterophily assumption through label domain analysis. The framework comprises three key components: a hypergraph-oriented label propagation method for deriving class-specific label features, a label tensor construction approach characterizing node-level heterophily intensity via 2D tensors, and a center attention mechanism that dynamically optimizes hypergraph structures. By enabling nodes to dynamically reconfigure the local graph structure based on microscopic heterophily intensity, HADHG effectively mitigates heterophily interference. Comprehensive experiments using real-flight data from Unmanned Aerial Vehicles and the public Gear dataset highlight the framework’s superiority over state-of-the-art methods. The codes and datasets are openly available at https://github.com/DL-LEO/HADHG.