Information integration for detecting communities in attributed graphs

Abstract

Real social networks can be described using two dimensions: first a structural dimension that contains the social graph, e.g. the actors and the relationships between them, and second a compositional dimension containing the actors' attributes, e.g. their profile. Each of these dimensions can be used independently to cluster the nodes and explain different phenomena occurring on the social network, whether from a connectivity or an individual perspective. In the case of community detection problem, an emergent research field explores how to include relationships and node attributes in an integrated clustering process. In this paper, we present a novel approach which integrate two partitions, one structural and one compositional, after they habe been generated by dedicated and specialized clustering steps. We rely on a contingency matrix with structural groups in rows and compositional ones in columns. The problem is to manipulate rows and columns to provide a new partition which maintains a good trade-off between both dimensions. In this paper we propose two strategies to control the combination. Tested on real-world social networks, the final partitions are evaluated in terms of entropy and density, and compared to pure structural or compositional partitions. The unified partitions show interesting properties, such as cohesive and homogeneous groups of actors. The method offers fine control on the combination process, giving new search capabilities to analysts without requiring the re-computation of the partitions.