Multi-view Semi-supervised Learning with a Unified Graph Convolutional Network

Abstract

In recent years, the proliferation of data-driven applications across diverse fields has sparked a surge in interest in applying semi-supervised learning to graphs. This surge is driven by the widespread use of graph data structures in real-world scenarios, such as interpersonal relationships in social networks, user behavior graphs in recommender systems, and molecular interaction networks in bioinformatics. However, certain data types, like images, pose challenges due to the lack of explicit graph structures and the presence of multiple view description methods. These complexities hinder the direct application of traditional semi-supervised learning techniques to graphs. As a result, researchers are exploring the integration of semi-supervised learning with deep learning to capitalize on its rich information and improve model performance. Effectively combining graph structures with multi-view data remains a daunting task that requires further investigation. This paper presents the Linear projection Fused Graph-based Semi-supervised Classification (LFGSC) method designed specifically for multi-view data, building upon the Graph Convolutional Network (GCN) architecture. Initially, for each view, we employ a semi-supervised approach to simultaneously estimate the corresponding graph and flexible linear data representations in a low-dimensional feature space. Subsequently, an adaptive and unified graph is constructed, followed by the use of a fully connected network to fuse the projected features and reduce dimensionality. Finally, the fused features and graph are fed into a GCN for semi-supervised classification. During training, the model incorporates cross-entropy loss, manifold regularization loss, graph auto-encoder loss, and supervised contrastive loss. Leveraging linear transformation significantly reduces the input feature dimensions for GCN, resulting in high accuracy while minimizing computational overhead. Additionally, experimental results on various benchmarked multiview image datasets demonstrate the superiority of LFGSC over existing semi-supervised learning methods for multi-view scenarios. Source code: https://github.com/BiJingjun/LFGSC.