关键词: 图卷积网络, 社区检测, 聚类, 模式识别

Abstract:

The graph convolutional neural network can effectively extract feature information on non-Euclidean distance data. This paper proposes an unsupervised community detection algorithm based on graph convolutional network model. The algorithm first selects some nodes in the graph to add artificial flags to simulate the input signal, satisfying the requirements of the propagation rule of the graph convolutional network. Then it passes the flags to the neighboring nodes through the modified graph convolutional network propagation rule. While comparing the different flags on the same node and getting the attribution result of the nodes, community partitioning matrix can be obtained by optimizing the result. Finally, it uses real-world data sets to test this algorithm and compares with some classic community detection algorithms. The experimental results show that the algorithm can achieve a better result in different types of graph data sets.

Key words: graph convolutional neural network, community detection, cluster, pattern recognition