基于NMF的社团及属性标签发现方法

Community and Attribute Label Detection Method Based on NMF

社团发现是社交网络分析的重要任务,有助于理解网络结构.目前基于网络拓扑的社团发现算法无法对发现的社团进行语义分析,例如社团的成因或属性等.拓扑结构和语义分析从两个角度对社团进行描述,但目前很少有研究同时分析这两个特性.本文提出基于联合非负矩阵分解的方法,将网络拓扑和节点内容信息(如节点属性等)结合起来,同时挖掘社团及其属性标签.该方法通过约束从拓扑和内容信息挖掘出的社团的相似性以及社团间关系的相似性,提高社团发现的准确率,并且从节点属性信息中挖掘社团的属性标签对社团进行语义分析.真实数据集上的实验结果表明,该方法能够有效地发现社团并对社团进行语义描述.

Community detection is one fundamental task in social network analysis which enables us to understand the structure of the network. Many community detection methods based on exploring network topologies have been developed,however,they cannot give certain semantic interpretation of the communities explored,such as why the members form a community,the attributes of a community and so on. Although topological structures and semantic interpretations provide complementary information from two perspectives for further understanding,little effort has been made to discover and analyze these two essential network properties together. By integrating network topology and node content information,e. g.,node attributes,we develop a method based on joint Nonnegative Matrix Factorization（ NMF）,in which the communities and attribute labels of them are derived simultaneously. Specially,both the communities and community relationships derived from topology and content information are pushed to be similar in the method,the community detection accuracy is improved consequently. Moreover,the attribute labels of the resultant communities are explored from node content information for semantic annotation. Experimental results on real-world networks show that our method can effectively find out the communities and semantically interpret the communities.