<p>Aligning complementary information across diverse views serves as the fundamental bedrock for extracting resilient representations in multi-view graph learning. However, existing architectures often overlook two limitations for consistency. First, existing paradigms fall short in harmonizing global cross-view regularities with the topology-aware characteristics inherent in local neighborhoods. This oversight leads to compromised node representations that lose discriminative details, failing to capture a node’s high-level role and subtle characteristics of its neighborhood. Second, the learning process often lacks explicit class-aware guidance, which causes the learned representations of different classes to overlap, compromising the final classification performance. In this study, we propose an end-to-end multi-view learning framework called Consensus-Prototype Guided Graph Convolutional Network (<b>CPG-GCN</b>). Essentially, CPG-GCN leverages a collaborative mechanism between structural adaptation and prototypical guidance to enhance multi-view representation learning. Specifically, CPG-GCN employs a Global-Local Consensus Regulator that dynamically balances the contributions from both cross-view consensus and neighborhood characteristics, ensuring optimal information integration for each node. In addition, CPG-GCN incorporates a Prototypical Guidance Constraint. By deploying learnable category anchors, this regularizer sculpts a highly discriminative latent space, compelling individual nodes to closely gravitate toward their respective class prototypes. The superiority of our proposed CPG-GCN is validated through experiments on eleven real-world datasets. The results show that our model achieves superior performance over state-of-the-art methods. Additionally, CPG-GCN demonstrates its strong robustness in low label-rate environments.</p>