Semi-supervised Learning of Graphs with Joint Node Similarity and Topology

In the context of semi-supervised learning, graph convolutional networks have become a hot research topic with their excellent ability to handle graph data. Existing graph convolutional network models have achieved success in classifying nodes on graphs, but they are also very limited by the preprocessing of the graph structure. Such networks face two problems: (1) the problem of confusing feature aggregation and node representation distortion due to too many dissimilar nodes around the nodes; (2) the problem of insufficient node representation learning due to the limited amount of information transmitted and received by the nodes with low degree. In this paper, we propose a joint node similarity and topology approach to solve these problems. We design a dual-view network to learn the similarity information of the nodes and the structural similarity information of the graph, and use interpolation perturbation in the Auxiliary View network to enhance the toughness of the model, and both of them perform joint learning by sharing parameters. Experimental results show that this method is more effective than traditional methods which only rely on structural information, and while achieving better classification accuracy on node classification, its learned node representation exhibits better inter-class separation and intra-class aggregation.