联合图的全局和局部互信息的重叠社团检测算法

图神经网络在学习节点表示中展现了其突出的能力，然而在社团检测方面，大多数图神经网络模型仍然使用K-means来定位社团中心，为了克服K-means不适用于高维空间下聚类的缺点，提出了联合图的全局和局部互信息的重叠社团检测算法（overlapping community detection algorithm using global and local mutual information of graph，overDGI），这是一种用于处理重叠社团检测问题的图神经网络。首先，采用最大化图互信息和社团互信息使得隶属于同一社团的节点间的向量表示距离更近、更接近社团中心；然后，设计了一个目标分布来帮助模型更好地解决重叠社团检测任务。综合实验表明，overDGI在重叠社团划分上的表现对比现有的几种基准算法都有很强的竞争力。

Overlapping community detection algorithm using global and local mutual information of graph

Graph neural network shows its outstanding ability in learning node representation. However, in terms of community detection, most graph neural network models still use K-means to locate community centers. In order to overcome the disadvantage that K-means is not suitable for clustering in high-dimensional space, this paper proposed an overlapping community of glo-bal and local mutual information of joint graphs, which was a graph neural network for dealing with overlapping community detection problems. First, it used the node vector representation to express the graph structure information more accurately by maximizing graph mutual information. Then, on the basis of using the graph structure to locate the community center, by maximizing the community mutual information, the vector representation distance between nodes belonging to the same community was closer and closer to the community center. Finally, it a designed target distribution to help the model solve the overlapping community detection task better. Through comprehensive experiments, it shows that overDGI has strong competitiveness compared with several existing benchmark algorithms in overlapping community division.