DAPF: a dual-view adaptive path framework for self-supervised heterogeneous graph representation learning

Abstract Heterogeneous information networks (HINs) contain rich type-aware semantics and complex relational structures, making effective representation learning challenging. Existing heterogeneous graph contrastive methods have explored cross-view alignment, but many construct semantic views through holistic meta-path or neighborhood aggregation, which may weaken the sequential context and instance-level semantic variations of concrete meta-path instances. To alleviate this issue, we propose DAPF, a Dual-View Adaptive Path Framework for self-supervised representation learning on HINs. DAPF jointly models multi-relational structural patterns and fine-grained meta-path instance semantics. Specifically, the Adaptive Structure View learns structure-aware node representations through adaptive aggregation over meta-path-induced structural views, while the Semantic Path View employs a Transformer-based path encoder with path-level attention to capture order-sensitive semantic dependencies within individual meta-path instances. A symmetric cross-view contrastive objective aligns the structural and semantic representations, while a prototype-level contrastive objective further promotes semantic organization in the latent space. Extensive experiments on ACM, DBLP, and AMiner over 10 independent representation-learning seeds demonstrate that DAPF achieves the strongest overall classification performance on ACM and DBLP. On AMiner, DAPF obtains the highest AUC and clustering performance, demonstrating strong ranking and unsupervised grouping capability. Additional ablation, component, sensitivity, robustness, and computational analyses further support the effectiveness of the proposed design and characterize its behavior under different experimental settings.