Labeling small-degree nodes promotes semi-supervised community detection on graph convolutional network

Community structure is one of the most important characteristics of network, which can reveal the internal organization structure of nodes. Many algorithms have been proposed to identify community structures in networks. However, the classification accuracy of existing unsupervised community detection algorithms is generally low. Therefore, the semi-supervised community detection algorithm which can greatly improve the classification accuracy by introducing a small number of labeled nodes has attracted much attention. Nevertheless, previous studies were sketchy in terms of label rates and also ignored the variation of classification accuracy under different labeling strategies. In this paper, based on graph convolutional networks, we first study the effect of labeling strategies and label rates on classification accuracy in four real world networks in detail. The research phenomenon is counter-intuitive but surprisingly effective: the classification accuracy of labeling small-degree nodes or random-selection nodes is significantly higher than that of labeling high-degree nodes. The labeling strategies based on acquaintance immune algorithm also prove this result. The interesting question that arises is what topological properties of the network can lead to such results? So we test and verify it in two kinds of synthetic networks. It is found that the phenomenon which labeling small-degree nodes promotes classification accuracy can be observed when the degree distribution of the network follows power-law distribution and the ratio of the external edges of the community to the total edges of nodes in the network is small. Graphical abstract