<p>Unsupervised graph representation learning with GNNs is critically important due to the difficulty of obtaining graph labels in many real applications. Recently, Graph Contrastive Learning (GCL) has emerged as a prominent approach for unsupervised graph learning, demonstrating remarkable performance across various tasks. However, existing GCL methods primarily focus on homogeneous or static heterogeneous graphs, leaving the dynamic heterogeneous graphs (DHGs) representation learning underexplored, which involve evolving node features and structures over time. In this paper, we propose a general unsupervised Dynamic Heterogeneous Graph Contrastive Learning (DHGCL) framework to address the challenges of capturing both heterogeneous structural information and temporal dynamics in DHGs. Specifically, we design novel contrastive learning tasks that leverage the intrinsic characteristics of DHG data. The temporal proximity view sampling layer is next designed based on the temporal proximity similarity to construct temporal augmented views, enabling effective modeling of temporal evolution. Additionally, a heterogeneous GNN is utilized to capture heterogeneity within static snapshots, and a cross-time aggregation module is used to integrate node embeddings across time spans. We conduct experiments on four real-world datasets, and the results demonstrate that our method outperforms numerous baseline methods, achieving at least 10% improvement on various datasets.</p>