Pseudo Labeling Collaborative Embedding Representation: A Multi-Channel GCNs Framework for Semi-Supervised Node Classification.

In recent years, Graph Convolutional Networks (GCNs) have emerged as a crucial methodology for handling graph-structured data, exhibiting superior performance in semi-supervised classification tasks. However, most existing GCNs encounter two main issues in real-world scenarios: (1) graph-structured data may be incomplete, for example, containing outlier nodes and noisy edges, which poses a great challenge for GCNs to extract the relation information for classification tasks; (2) the scarcity of labeled data, often limited to a few-shot scenario, hampers the ability of GCNs to learn comprehensive embedding representations. To cope with these issues, we propose a novel framework called pseudo labeling collaborative multi-channel graph convolutional networks (PCM-GCN). First, considering incomplete graph-structured data, we develop two modules: graph generation module and multi-channel fusion module. The graph generation module is designed to extend the raw data to multiple graphs, which avoids being constrained by the effective expression ability of the raw data. Meanwhile, the multi-channel fusion module integrates embeddings from multiple graphs, capturing the complementarity among multiple channels. Second, to address the problem of sparse labels, we develop a confidence-based pseudo labeling module, appending confident data with pseudo label to the labeled set to enlarge the training set. PCM-GCN leverages pseudo labeling to enhance multi-channel embedding fusion, resulting in rich and comprehensive node embedding representation. Extensive experiments on five benchmark datasets have shown that PCM-GCN surpasses other state-of-the-art methods in semi-supervised node classification tasks.