K阶图卷积属性网络社团检测方法

Method of K-order Graph Convolution Attribute Network Community Detection

挖掘属性网络中的社团结构有助于对网络节点进一步分析,具有重要的现实意义。图卷积神经网络能够有效地将属性网络的结构信息进行嵌入,获取节点的特征表示,从而可获得性能良好的社团结构。然而,现有图卷积方法大多使用固定的低阶图卷积,只考虑每个节点一阶或二阶内的邻居,没有充分利用节点关系,忽略了网络结构的多样性。另外原始网络结构的稀疏性无法克服,会降低社团检测的性能。为解决上述问题,提出一种融合属性信息与结构信息的K阶图卷积社团检测方法(KGCN),该方法可以有效地克服原始网络的稀疏性并利用节点的高阶结构进行社团检测。首先根据节点的属性信息对原始网络进行重构,缓解原始网络结构的稀疏性;其次考虑到高阶结构关联,采用K阶图卷积编码器对节点进行编码,获得节点的特征表示;最后使用谱聚类算法进行社团检测。实验结果表明,在四个真实数据集上,相比现有算法,KGCN方法取得更好的社团检测结果。

Mining community structure in attribute network is helpful to further analysis of network nodes,which has important practical significance.Graph convolution neural network can effectively embed the structural information of attribute network and obtain the feature representation of nodes.Based on this,the community structure with good performance can be obtained.However,most of the existing graph convolution methods use fixed low-order graph convolution,only consider the first-order or second-order neighbors of each node,don’t make full use of the node relationship and ignore the diversity of network structure.In addition,the sparsity of the original network structure can’t be overcome,which will reduce the performance of community detection.In order to solve the above problems,this paper proposes a K-order graph convolution community detection method(KGCN)which combines attribute information and structure information.This method can effectively overcome the sparsity of the original network and use the high-order structure of nodes for community detection.Firstly,the original network is reconstructed according to the attribute information of nodes to alleviate the sparsity of the original network.Secondly,considering the high-order structure correlation,the K-order graph convolution encoder is used to encode the nodes to obtain the feature representation of the nodes.Finally,community detection is carried out using spectral clustering algorithm.Experimental results show that the KGCN method achieves better community detection results than the existing algorithms on four real datasets.