Semi-supervised node classification via graph learning convolutional neural network

Graph convolutional neural networks (GCNs) have become increasingly popular in recent times due to the emerging graph data in scenes such as social networks and recommendation systems. However, engineering graph data are often noisy and incomplete or even unavailable, making it challenging or impossible to implement the de facto GCNs method directly on them. Current efforts for tackling this issue either require an overparameterized model that is hard to scale, or simply re-weight the existing edges for different downward tasks. In this work, we tackle this problem through introducing a graph learning convolutional neural network (GLCNN), which can be employed on both Euclidean space data and non-Euclidean space data. The similarity matrix is learned by a supervised method in the graph learning layer of the GLCNN. Moreover, graph pooling and distilling operations are utilized to reduce over-fitting. Comparative experiments are done on three different datasets: citation dataset, knowledge graph dataset, and image dataset. Results demonstrate that the GLCNN can improve the accuracy of the semi-supervised node classification by mining useful relationships among nodes. The performance is more obvious especially on datasets of Euclidean space. Specifically, GLCNN outperforms the best baseline by 3.1% and 1.1% on MNIST and SVHN datasets. Moreover, the robustness is explored by adding noises on the edge of the graph. Sensitive analysis and visualizations are performed to demonstrate effects of some key parameters.