Abstract: To address the limitation of existing dynamic graph anomaly detection methods in accurately modeling anomaly patterns at different time scales, a method of spatio-temporal contrastive learning for dynamic graph anomaly detection (STCL-DGAD) is proposed. The proposed method consists of three modules: spatio-temporal feature extraction, spatio-temporal contrastive learning, and dynamic anomaly detection. First, the spatio-temporal feature extraction module employed time encoding and a Transformer-based architecture to capture node features incorporating temporal dynamics and global structural information. Second, in the spatial-temporal contrastive learning module, a global-local dualgranularity contrastive mechanism was designed. The global contrastive part aggregated historical interaction information throuhg exponentially decaying weights, and further constructed an augmented graph to model long-term spatial-temporal dependencies. The local contrastive part captured short-term abnormal mutations by aligning features of adjacent time snapshots. Finally, the dynamic anomaly detection module enhanced the model robustness via a regularization strategy. Experimental results on four real-world datasets demonstrate that STCL-DGAD outperforms existing methods in anomaly detection tasks.