DualCL: Dual contrastive learning for hypergraphs

Unsupervised hypergraph representation learning is often restricted to a single structural perspective, with hyperedges serving mainly as intermediaries for node-centric message passing. To address this limitation, we propose DualCL, a dual contrastive learning framework that exploits the intrinsic duality of hypergraphs. DualCL constructs the original hypergraph and its dual as two endogenous views of the same incidence system, thereby modeling high-order relations from the node and hyperedge domains, respectively. It then aligns the two views at both the node and hyperedge levels: an original node is paired with its corresponding dual hyperedge, while an original hyperedge is paired with its corresponding dual node. A sampled node–hyperedge incidence reconstruction objective is further introduced to preserve the original high-order topology. Experiments on five public hypergraph datasets show that DualCL achieves competitive performance in node classification and unsupervised clustering. Ablation studies verify the contributions of the dual-view design, the two contrastive objectives, and incidence reconstruction, while scalability experiments demonstrate the effectiveness of the sampled training objectives on large sparse hypergraphs.