计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (11): 73-83.DOI: 10.3778/j.issn.1002-8331.2110-0487

• 大数据与云计算 • 上一篇    下一篇

图卷积网络增强的非负矩阵分解社区发现方法

郑裕龙,陈启买,贺超波,刘海,张晓雨   

  1. 1.华南师范大学 计算机学院,广州 510631
    2.仲恺农业工程学院 信息科学与技术学院,广州 510225
  • 出版日期:2022-06-01 发布日期:2022-06-01

Nonnegative Matrix Factorization Community Detection Method  Enhanced by Graph Convolutional Network

ZHENG Yulong, CHEN Qimai, HE Chaobo, LIU Hai, ZHANG Xiaoyu   

  1. 1.School of Computer Science, South China Normal University, Guangzhou 510631, China
    2.School of Information Science and Technology, Zhongkai University of Agriculture and Engineering, Guangzhou 510225, China
  • Online:2022-06-01 Published:2022-06-01

摘要: 非负矩阵分解(nonnegative matrix factorization,NMF)因其有效性和易解释性强被广泛应用于社区发现领域。然而,现有大多数基于NMF的社区发现方法都是线性的,无法有效处理复杂网络的非线性特征,从而导致社区发现性能还有待进一步提高。针对该问题,提出了一种图卷积网络(graph convolutional network,GCN)增强的非线性NMF社区发现方法NMFGCN。NMFGCN包含两个主要模块:GCN和NMF,其中GCN用于学习网络节点表示,NMF将节点表示作为输入获得网络的社区表示。此外,提出一个联合优化方法以训练NMFGCN,不仅使得NMFGCN具有非线性特征表示能力,而且可以使得GCN和NMF相互促进并获得更好的社区划分结果。在人工合成网络和真实网络上进行大量实验,结果表明NMFGCN优于目前基于NMF的社区发现方法,从而证明NMFGCN确实可以提高NMF社区发现方法的性能。此外,NMFGCN还优于DeepWalk和LINE常用图表示学习方法。

关键词: 社区发现, 非负矩阵分解, 图卷积网络, 非线性方法

Abstract: Nonnegative matrix factorization(NMF) has been widely used in community detection due to its effectiveness and better interpretability. However, most existing NMF-based methods are linear, and consequently cannot effectively process the nonlinear characteristics of complex networks. As a result, the performance of community detection needs to be further improved. In view of this, NMFGCN is proposed, which is a nonlinear NMF-based method enhanced by graph convolutional network(GCN). NMFGCN contains two main modules:GCN and NMF. GCN is applied to obtain the representation of nodes, where NMF takes the representation as input to generate community representation of a given network. In addition, a joint optimization method is proposed to train NMFGCN, which not only allows NMFGCN to possess nonlinear trait, but also enables its two modules to be optimized jointly. Extensive experiments are conducted on synthetic networks and real networks. The results show that NMFGCN is superior to state-of-the-art NMF-based methods, and thereby proves that NMFGCN indeed can boost the performances of NMF-based methods. Moreover, NMFGCN also outperforms some widely-used graph representation based methods, such as LINE and Deepwalk.

Key words: community detection, nonnegative matrix factorization, graph convolutional network, nonlinear method