Compositional Subgroup Discovery on Attributed Social Interaction Networks.

While standard methods for detecting subgroups on plain social networks focus on the network structure, attributed social networks allow compositional analysis, i. e., by exploiting attributive information. Accordingly, this paper applies a compositional perspective for identifying compositional subgroup patterns. In contrast to typical approaches for community detection and graph clustering it focuses on the dyadic structure of social interaction networks. For that, we adapt principles of subgroup discovery – a general data mining technique for the identification of local patterns – to the dyadic network setting. We focus on social interaction networks, where we specifically consider properties of those social interactions, i. e., duration and frequency. In particular, we present novel quality functions for estimating the interestingness of a subgroup and discuss their properties. Furthermore, we demonstrate the efficacy of the approach using two real-world datasets on face-to-face interactions.