Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning

TLDR

Graph convolutional networks (GCNs) have emerged as a promising approach for graph‑based semi‑supervised learning, integrating vertex features and topology, yet their internal mechanisms remain unclear and they still require substantial labeled data for validation. The study aims to deepen understanding of GCNs and overcome their limitations by proposing co‑training and self‑training strategies. The authors analyze GCNs as Laplacian smoothing, highlight over‑smoothing with many layers, and introduce co‑training and self‑training methods that allow training with very few labels without extra validation data. Experiments on benchmark datasets confirm that the proposed co‑training and self‑training methods substantially improve GCN performance with few labels and eliminate the need for additional validation labels.

Abstract

Many interesting problems in machine learning are being revisited with new deep learning tools. For graph-based semi-supervised learning, a recent important development is graph convolutional networks (GCNs), which nicely integrate local vertex features and graph topology in the convolutional layers. Although the GCN model compares favorably with other state-of-the-art methods, its mechanisms are not clear and it still requires considerable amount of labeled data for validation and model selection. In this paper, we develop deeper insights into the GCN model and address its fundamental limits. First, we show that the graph convolution of the GCN model is actually a special form of Laplacian smoothing, which is the key reason why GCNs work, but it also brings potential concerns of over-smoothing with many convolutional layers. Second, to overcome the limits of the GCN model with shallow architectures, we propose both co-training and self-training approaches to train GCNs. Our approaches significantly improve GCNs in learning with very few labels, and exempt them from requiring additional labels for validation. Extensive experiments on benchmarks have verified our theory and proposals.