Topology-preserving and structure-aware (hyper)graph contrastive learning for node classification

Abstract

Recently, graph contrastive learning (GCL) has attracted considerable attention, establishing a new paradigm for learning graph representations in the absence of human annotations. While notable advancements have been made, simultaneous consideration of both graphs and hypergraphs remains rare. This limitation arises because graphs and hypergraphs encode connectivity differently, making it challenging to develop a unified structure augmentation strategy. Conventional structure augmentation methods like adding or removing edges risk imperiling intrinsic topological traits and introducing adverse distortions such as disconnected subgraphs or isolated nodes. In this work, we propose a framework of contrastive learning on graphs and hypergraphs, named as UniGCL, to address these challenges by leveraging a unified adjacency representation that enables simultaneous modeling of pairwise and higher-order relationships. In particular, two structure augmentation methods are developed to perturb graph structure weights instead of altering connectivity, thereby preserving both graph and hypergraph topology while generating diverse augmented views. Furthermore, a structure-aware contrastive loss is proposed, which incorporates gradient perturbation techniques to enhance the model’s ability to capture fine-grained structural dependencies in (hyper)graphs. Extensive experiments are conducted on six real-world graph datasets and nine representative hypergraph datasets to evaluate the performance of the proposed framework. The results demonstrate that UniGCL achieves superior node classification performance compared to the advanced graph and hypergraph contrastive learning methods, across datasets with different homophilic extents and limited annotations. Additionally, ablation studies validate the effectiveness of our structure-preserving augmentations and structure-aware contrastive loss in enhancing performance.