A Graph Convolution model considering label co-occurrence LC_GCN

Graph convolution network (GCN) is a Semi-supervised algorithm that applies the idea of convolution to graph structure data, and it is used for node classification tasks in graphs. Original algorithm only considers the characteristics and adjacency of the nodes in the graph, but fails to consider the association between label and simply represents the label as a one-hot vector. In this paper, we propose LC_GCN. This model contains a label convolution module based on the original GCN, and use it to get a better classifier. An open pre-trained word vector is used as the label feature, and we designed an algorithm to use the conditional probability of the association between label to generate adjacency matrix of labels to obtain the classifier by GCN. Then combine it with the original node GCN, and put the vector obtained by the node through the GCN into this classifier. Experimental results show that our proposed LC_GCN outperforms the existing algorithms.