<p>Heterogeneous graph neural networks (HGNNs) have showcased exceptional modeling prowess in characterizing intricate structures and diverse semantic information. Despite notable advancements in capturing high-order relational patterns among heterogeneous nodes, existing methods often falter in accurately expressing local neighborhood information, resulting in constraints on global–local structural modeling. To address this, we introduce a Heterogeneous graph neural network with Multi-scale Meta-path Contrastive learning (HMMC), a self-supervised framework designed to enhance representation learning in heterogeneous graphs. Our approach incorporates a multi-scale meta-path embedding mechanism that concurrently captures local and global structural information, thereby alleviating the challenge where overly short meta-paths yield limited representation while overly long meta-paths introduce noise. Furthermore, we devise a cross-view self-supervised contrastive learning framework that bolsters the model’s capacity to learn heterogeneous graph structures by executing contrastive optimization across various perspectives, enabling more precise modeling of intricate node relationships. To address the issue of noisy negative samples in traditional contrastive learning—which often hinders model optimization—we propose a star-shaped contrastive loss function. This inventive loss function exploits center-neighborhood structural dependencies as supervisory signals for self-contrastive learning, ensuring representation consistency among positive pairs while circumventing over-smoothing caused by the convergence of similar node embeddings. By integrating these two highly synergistic modules, our method resolves the inherent trade-off between local structural granularity and global semantic consistency in heterogeneous graph representation learning. To validate the effectiveness of HMMC, we conduct comprehensive experiments on multiple public heterogeneous graph datasets. The results demonstrate that HMMC surpasses state-of-the-art baselines, achieving performance gains of 0.5%–4.1% across different benchmarks. These findings substantiate the superiority of HMMC in terms of representation power, robustness, and generalization capability for heterogeneous graph learning tasks.</p>