Graph clustering assigns nodes into tightly connected groups known as communities, and community detection algorithms traditionally focus on non-attributed networks that only provide a partial representation of the underlying systems. Thus, community detection in attributed networks becomes a hot topic since it provides an insight into the structure-function relation of the underlying systems. However, identifying community in attributed networks is highly non-trivial since it simultaneously takes into consideration both topological structure and node attributes. Current algorithms detect communities by either incorporating attributes into structure of networks or directly fusing features of nodes and structure of networks, which cannot effectively handle the heterogeneity of structure and attributes, hampering fully exploitation of them. To overcome these problems, we propose a novel algorithm by joint nonnegative matrix factorization and graph optimization (called NMFjGO) for community detection in attributed networks. To overcome the heterogeneity of structure and attributes, an attribute similarity matrix for nodes is constructed in terms of attribute profiles. To obtain the latent feature of attributed networks for community, nonnegative matrix factorization (NMF) is adopted to jointly factorize the established attribute similarity matrix and adjacent matrix of networks. The latent feature of node attributes are dynamically fused with the topological structure of attributed networks. The advantage of NMFjGO is that it jumps out of local minima by dynamically exploring the attribute and links of networks. The experimental results demonstrate that the proposed algorithm significantly outperforms state-of-the-art algorithms in terms of accuracy on a number of attributed networks.

图集群将节点分配到紧密相连的组（称为社区）中，而社区检测算法传统上专注于仅提供底层系统的部分表示的非属性网络。因此，属性网络中的社区检测成为热门话题，因为它提供了对底层系统的结构-功能关系的深入了解。但是，在归属网络中识别社区非常重要，因为它同时考虑了拓扑结构和节点属性。当前的算法通过将属性合并到网络结构中或直接融合节点和网络结构的特征来检测社区，这无法有效处理结构和属性的异质性，从而阻碍了对它们的充分利用。NMFjGO），用于归属网络中的社区检测。为了克服结构和属性的异质性，根据属性配置文件构造了节点的属性相似性矩阵。为了获得社区属性网络的潜在特征，采用非负矩阵分解（NMF）共同分解建立的属性相似度矩阵和网络的相邻矩阵。节点属性的潜在特征与属性网络的拓扑结构动态融合。NMFjGO的优势它是通过动态探索网络的属性和链接而跳出局部最小值的。实验结果表明，该算法在许多属性网络上的准确性方面明显优于最新算法。