Breaking the curse of dimensional collapse in graph contrastive learning: A whitening perspective

Yang Tao, Kai Guo, Yizhen Zheng, Shirui Pan, Xiaofeng Cao, Yi Chang

Abstract: Highlights•The paper provides an intuitive analysis, indicating the dimensional collapse in GCL is inevitable.•The paper introduces a novel plug-and-play module (WGCL), effectively addressing the dimensional collapse issue.•WGCL module significantly improves the performance of GCL backbones on various datasets.