MHGCL: Combining Motif and Homogeneity in graph contrastive learning

Chunying Li, Yanxi Guo, Zhikang Tang, Wenbo Li, Yang Chen, Jinli Cao, Yanchun Zhang, Jianjian Jiang

Abstract: Contrastive learning, as a paradigm for learning without relying on labeled data, has become one of the mainstream research hotspots in graph clustering. Homophily is key to improving contrastive learning performance, yet the importance of higher-order structures, which are vital for graph-based structures, has often been overlooked. This paper introduces the Motif and Homogeneity based Graph Contrastive Learning (MHGCL) approach, integrating higher-order graph characteristics and graph homophily within a single framework. Initially, the method leverages the triangular motif structure within graph networks to extract higher-order structural features, effectively capturing complex relationships. Further, based on the consideration of graph homophily, we use the k-means algorithm to obtain centroids and then combine the Gaussian Mixture Models (GMM) to calculate and consider both the posterior and prior probabilities of nodes. Thereafter, we obtain a clustering assignment matrix and the true affinity relationships between nodes. By categorizing in-class and inter-class positive and negative samples based on the affinity relationships, we effectively expand the positive samples in contrastive learning. Comprehensive experiments show that our MHGCL method achieve significant performance improvements across different Graph Contrastive Learning (GCL) tasks.

External IDs:dblp:journals/www/LiGTLCCZJ25