Modified partition integration method for community detection in multidimensional social networks

Abstract

A social network may be viewed as a network of interactions or relationships amongst nodes that group together to form communities. In the real world, social networks involve multiple types of interactions that may evolve over time and, therefore, are inherently multidimensional. Extracting a community structure that best represents different dimensions in a multidimensional social network is a challenging problem. Tang et al. (2012) proposed a partition integration method for detection of communities in a d-dimension social network. It starts by determining hard clustering of the individual dimensions using the k-means algorithm, followed by application of cluster ensemble to obtain final clustering. However, the k-means algorithm is not suitable here because of its high sensitivity to the initial condition that is likely to yield results with relatively high variance. In this paper, we propose a variant of Tang's method, using state-of-the-art unidimensional community detection algorithms, both evolutionary and non-evolutionary, instead of k-means, for hard clustering. We measure the quality of the final partitioning by calculating the average overlap between the final and dimension-wise partitions. The proposed algorithm successfully detects the community structure for real as well as synthetic multidimensional networks, outperforming the state-of-the-art algorithms. The proposed method works well for both dynamic and static multidimensional networks and a priori knowledge of the number of communities is not required. It works well for both weighted and unweighted undirected networks, is easily extendable for directed networks, and does not require any additional decoding of the result.