Cross-scale hypergraph contrastive clustering

Abstract Hypergraphs have demonstrated remarkable advantages in numerous fields due to their capability to model multi-way correlations among an arbitrary number of vertices in complex data. However, existing contrastive learning-based methods face two core challenges in unsupervised hypergraph clustering. First, the uniform random sampling strategy is prone to introducing negative sampling bias under the message passing mechanism of hypergraphs. Second, the absence of explicit clustering-guided constraints results in a disconnection between representation learning and the clustering objective. To tackle these issues, a Cross-scale Hypergraph Contrastive Clustering (CHCC) framework is proposed in this paper. CHCC constructs a hierarchical discriminative system comprising “micro-scale nodes, meso-scale hyperedges, and macro-scale cluster prototypes.” Specifically, via a node-level topological discrimination mechanism, this framework extends the contrastive objective from a “node-to-node” paradigm to a “node-to-hyperedge” one, fundamentally circumventing the sampling bias induced by false negatives. Simultaneously, a clustering-guided hyperedge semantic discrimination mechanism is introduced to achieve clustering-oriented feature optimization by contrasting local structural semantics with global cluster prototypes. Experimental results on six benchmark datasets demonstrate that CHCC outperforms existing state-of-the-art methods across multiple clustering metrics, comprehensively verifying its effectiveness in synergistically optimizing topology preservation and clustering performance.