Heterogeneous Graph Contrastive Masked Autoencoder

Contrastive Learning (CL) and Masked Autoencoder (MAE) methods have been widely applied to Self-Supervised Learning (SSL) for Heterogeneous Graphs (HGs). However, existing graph MAE methods mainly focus on reconstructing node features and topological structures, while lacking the ability to perceive high-order semantic information, which limits the discriminative power of the learned representations. To address this issue, this paper proposes a novel SSL framework, termed Heterogeneous Graph Contrastive Masked Autoencoder (HGCMAE), to enhance the representational discriminability of traditional MAE-based methods. HGCMAE consists of four core modules, namely a hierarchical masking module, a generative branch, a contrastive branch, and a joint optimization module. Specifically, the hierarchical masking module applies controllable multi-level masking to node attributes and meta-path-based adjacency matrices, thereby simulating realistic data incompleteness while preserving sufficient local cues for reconstruction. Based on these masked views, the generative branch reconstructs masked node attributes and meta-path-induced subgraphs from adaptive semantic embeddings, thereby preserving relation-aware semantic feature interactions. Meanwhile, the contrastive branch leverages coherent structural embeddings to enforce cross-view consistency under topology-aware structural dependencies across meta-path-induced graphs. Finally, the joint optimization module jointly constrains the two branches under the guidance of cross-view high-order semantic correlations and dynamically balances the generative and contrastive objectives. Experimental results on four real-world HG datasets demonstrate that HGCMAE achieves competitive performance in node classification and clustering tasks.