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