面向带属性复杂网络的鲁棒,强解释性社团发现方法

Semantic community identification contains two aspects, i.e., to improve the accuracy of community detection and to annotate the semantics of communities precisely. Traditional community detection methods always assume that the network topology and node contents share the same community memberships. However, this assumption does not always hold in many real-world networks. For example, in a Twitter network, social links usually directly reflect which users gather into a community, while users may generate diverse and disordered content information. To solve this problem, it is necessary to extract useful content information to assist topology information in detecting out more actual and accurate communities. In this paper, we carefully rethink the relationships between community structure, topic cluster, network topology and node contents, and propose a novel generative model different from the traditional generative model. In the new generative model, we logistically give a more reasonable explanation about the relationships between community structure, topic cluster, network topology and node contents. Under the drive of the new generative model, we design a new community detection method, referred to as Robust and Strong Explanatory Community Detection (RSECD). To be specific, based on nonnegative matrix factorization (NMF), we are able to obtain the community membership matrix for network topology and cluster membership matrix for node contents. More importantly, there exists some latent relationship between network communities and content clusters (with semantics), thus we innovatively introduce a transition matrix with a suitable prior to describe this relationship. As a result, even though the content information does not match with topology information, our method can still obtain accurate detection results by using the transition matrix with a suitable prior. At last, we put network topology, node content and transition matrix into a unified NMF framework and optimize them altogether by designing effective updating rules in order to achieve an integral balance of them. Furthermore, we analyze RSECD's calculational complexity after taking into account the sparsity of the adjacency matrix and attribute matrix. To justify our approach's effectiveness and robustness, we conduct extensive experiments. Firstly, we employ the Newman's model to generate artificial benchmark networks, then we use the generated networks to analyze the parameter in our objective function and verify that RSECD can solve topology and content's mismatch problem in the network well. The results of artificial networks experiment show our approach's strengths, i.e., effectiveness and robustness. Next, we compare our method with eight state-of-the-art community detection approaches on seven real-world networks. The experimental results show that RSECD achieves up to 6%-14% lift in comparison with the best baseline algorithm under four different kinds of community detection metrics, which further demonstrate RSECD's superior performance. We also report the running time of RSECD and other baseline algorithms spending on seven real-world networks. The results show that RSECD's time cost is far less than the average running time of other baselines on all real-world networks. Finally, in order to validate RSECD's strong interpretability to detected communities, we use a case study on a musical social network to semantically explain the hidden meanings of some topics and tell the 'true stories' behind communities.