InfoHarmonizer Graph Contrastive Clustering

Abstract

In graph deep clustering, the graph Laplacian filter is a widely used graph signal preprocessing tool that effectively smooths feature information. However, it cannot simultaneously aggregate graph structural information while smoothing features. Moreover, improper use of the graph Laplacian filter may lead to over-smoothing of feature information, adversely affecting clustering performance. To address these issues, we propose the InfoHarmonizer Graph Contrastive Clustering (IGCC) model. The model primarily comprises the Graph InfoHarmonizer preprocessing module and a structural aggregation contrastive network. Our designed Graph InfoHarmonizer module effectively aggregates graph structural information and smooths feature information. Notably, we use the KNN algorithm to aggregate feature neighbors to obtain graph structural information. We designed a simple graph structure contrastive learning network, which does not require complex data augmentation and uses graph diffusion to reduce the occurrence of feature over-smoothing. Experiments on five public datasets demonstrate the superiority of the IGCC model.