Graph Representation Learning with a Multi-View Hybrid Model for Class-Imbalanced Node Classification

Prevalent techniques for class-imbalanced graph learning predominantly rely on synthetic data generation to augment training samples for Graph Neural Networks (GNNs). Despite their popularity, these generative paradigms often incur substantial computational costs, introduce artifacts that distort intrinsic node relationships, and may destabilize model convergence. To address these limitations, we propose GraphMVHM (Graph Representation Learning with a Multi-View Hybrid Model), a versatile and efficient framework for node classification on imbalanced graph data. Specifically, the model constructs three complementary views: an original view, a generative view that supplements minority class samples, and a semantic view that captures node-topology correlations. Through cross-view collaborative learning, the framework mitigates the dominance of majority classes over minority classes, yielding a more discriminative node representation set compared to prior methods. The semantic view is constructed by integrating node features with topo-logical attributes to capture fine-grained semantic information at the node level, enabling a deeper understanding of each node’s intrinsic properties in relation to its neighbor-hood structure. Extensive experiments conducted on eight datasets—including three balanced, three imbalanced, and two large-scale graph benchmarks—demonstrate that the pro-posed method achieves state-of-the-art performance. Notably, on the CiteSeer dataset, accuracy improvements of 7% are achieved with both GCN and GAT backbones; on the Photo and Computers datasets, accuracy gains of 6% and 8% are attained with GAT, respectively.