Enhancing Semi-Supervised Multi-View Graph Convolutional Networks via Supervised Contrastive Learning and Self-Training

The advent of graph convolutional network (GCN)-based multi-view learning provides a powerful framework for integrating structural information from heterogeneous views, enabling effective modeling of complex multi-view data. However, existing methods often fail to fully exploit the complementary information across views, leading to suboptimal feature representations and limited performance. To address this, we propose MV-SupGCN, a semi-supervised GCN model that integrates several complementary components with clear motivations and mutual reinforcement. First, to better capture discriminative features and improve model generalization, we design a joint loss function that combines Cross-Entropy loss with Supervised Contrastive loss, encouraging the model to simultaneously minimize intra-class variance and maximize inter-class separability in the latent space. Second, recognizing the instability and incompleteness of single graph construction methods, we combine both KNN-based and semi-supervised graph construction approaches on each view, thereby enhancing the robustness of the data structure representation and reducing generalization error. Third, to effectively utilize abundant unlabeled data and enhance semantic alignment across multiple views, we propose a unified framework that integrates contrastive learning in order to enforce consistency among multi-view embeddings and capture meaningful inter-view relationships, together with pseudo-labeling, which provides additional supervision applied to both the cross-entropy and contrastive loss functions to enhance model generalization. Extensive experiments demonstrate that MV-SupGCN consistently surpasses state-of-the-art methods across multiple benchmarks, validating the effectiveness of our integrated approach. The source code is available at https://github.com/HuaiyuanXiao/MVSupGCN

💡 Research Summary

This paper proposes MV-SupGCN, a novel and comprehensive framework designed to address key limitations in semi-supervised learning for multi-view graph data. The core challenge lies in effectively leveraging limited labeled data alongside abundant unlabeled samples across multiple, complementary views or feature sets.

The authors identify three primary shortcomings in existing methods: 1) insufficient constraints for learning compact intra-class and separable inter-class representations in Graph Convolutional Network (GCN) models, 2) reliance on unstable or incomplete single-graph constructions per view, and 3) underutilization of unlabeled data, often limited to simple pseudo-labeling without exploring their underlying structural semantics.

To overcome these issues, MV-SupGCN integrates three synergistic components. First, it introduces a joint loss function that combines the standard Cross-Entropy loss with a Supervised Contrastive (SupCon) loss. The SupCon loss explicitly pulls embeddings of samples from the same class together while pushing apart embeddings from different classes in the latent space, thereby enhancing the discriminative power and generalization ability of the learned features.

Second, to build more robust and informative graph structures, the model employs a dual-graph construction strategy for each view. It simultaneously creates a K-Nearest Neighbor (KNN) graph, which captures local geometric structure, and a semi-supervised graph, which infers global semantic similarities by propagating the limited label information. The adjacency matrices from these two complementary graphs are fused, providing a richer and more stable structural representation for the GCN to process.

Third, the framework proposes a unified learning strategy that deeply exploits unlabeled data by integrating pseudo-labeling into the contrastive learning paradigm. High-confidence pseudo-labels generated by the model are used as supervisory signals for unlabeled samples. These signals are then incorporated into both the classification (cross-entropy) and the representation learning (supervised contrastive loss) objectives. Specifically, contrastive learning is applied to align the embeddings of the same unlabeled instance from different views, enforcing semantic consistency across views and leveraging the structural information within the unlabeled set.

Extensive experiments on seven public multi-view datasets demonstrate that MV-SupGCN consistently outperforms state-of-the-art graph-based multi-view learning methods. Ablation studies further confirm the individual contribution and necessity of each proposed component—the joint loss, the dual-graph construction, and the pseudo-labeling-enhanced contrastive learning. The results validate that the integrated approach successfully tackles the identified research gaps, leading to superior classification accuracy and robust feature representations. The paper concludes by discussing the model’s limitations and suggesting future directions, such as dynamic graph construction and scalability improvements.