Contrastive learning of graphs under label noise

Xianxian Li, Qiyu Li, De Li, Haodong Qian, Jinyan Wang

Abstract: Highlights•A new robust contrastive learning is proposed to mitigate labeling noise.•A dynamic cross-entropy loss is proposed to mitigate the overfitting.•The cross-space consistency is proposed as a bridge between the contrastive and dynamic cross-entropy loss.•Extensive experiments show the performance of our algorithm in resisting label noise.