LNAHC: Lightweight Neural Attributed Hypergraph Clustering via Multi-hop Attributed HNCut
Abstract
Attributed hypergraph clustering (AHC) is crucial for uncovering meaningful communities in complex relational data that exhibit both multi-way interactions and rich node attributes. While recent learning-based AHC methods have achieved notable advances, they suffer from three major limitations: (1) most methods decouple representation learning and clustering into separate stages, resulting in suboptimal non-cluster-related embeddings; (2) recent methods, such as CACH, introduce clustering-related losses but rely on heuristic alignment between soft cluster assignments and hard pseudo-labels, failing to optimize the underlying structure-attribute clustering objective; and (3) many approaches depend on computationally intensive contrastive learning and attention mechanisms, hindering scalability. To address these challenges, we propose LNAHC, a lightweight, end-to-end neural framework for AHC that comprises two key components: a novel attribute-aware hypergraph random walk strategy for constructing a structure-attribute proximity matrix that preserves both local and global semantic relationships, and a differentiable multi-hop attributed hypergraph normalized cut module that directly optimizes a relaxed cut-like loss, thereby regularizing the embedding space to produce high-quality clusters without post-processing or heavy architectural components. Extensive experiments on seven datasets demonstrate that the proposed LNAHC achieves superior performance compared to nine baselines.