Improving Graph Clustering with Multi-Granularity Debiased Contrastive Learning

Anonymous

Abstract: Recently, deep graph clustering achieves significant success by utilizing both the node attribute features and the graph structure information. However, the existing methods still have some limitations: (1) lack of a flexible mechanism to fuse multi-granularity information learned from different views. (2) introduce the noise positive-negative sample pairs lead to reduced the model performance. To tackle these problems, we propose a debiased contrastive learning framework DCL-MGI, which integrates the multi-granularity information of graph data. Specifically, two contrastive learning modules are constructed to capture multi-granularity feature information from node-level and graph-level, respectively. Meanwhile, an adaptive strategy of fusing stable graph structure information and node representations is proposed to select unbiased contrastive sample pairs, which reduces the false-negative samples. Furthermore, we utilize the temporal entropy metric to evaluate the sample quality under each view and communicate the two independent contrastive learning modules in a collaborative training manner. Experimental results on six real-world datasets demonstrate that our proposed framework enhances state-of-the-art methods on the graph clustering task.

Paper Type: long