Dual-View Contrastive Learning with Multi-Scale Label Propagation

Heterogeneous graphs contain diverse semantic information through multiple node and edge types, yet existing methods often tend to overlook the label information embedded in the graph structure. To overcome this issue, we introduce a dual-view contrastive learning model that integrates multi-scale label propagation and pseudo-supervision enhancement mechanisms (DCL-MS) for more robust node representation learning. The proposed model constructs two complementary views. One is a structure-aware view that employs multi-scale label propagation with pseudo-supervision filtering, while the other is a feature-learning view that integrates graph convolution and attention mechanisms. A cross-view contrastive loss is introduced to align the two views and enhance the consistency of learned representations. Results on two benchmark heterogeneous graphs indicate the effectiveness and stability of our method in both classification and clustering tasks.