Overlapping Communities Meet Roles and Respective Behavioral Patterns in Networks with Node Attributes

We present a new approach to the unsupervised and joint analysis of overlapping communities, roles and respective behavioral patterns in networks with node attributes. The proposed approach relies on an innovative Bayesian probabilistic model of the statistical relationships among communities, roles, behavioral role patterns and attributes. Essentially, under the devised model, behavioral role patterns define the abstract social functions underlying roles. Also, attributes are affiliated to roles. Moreover, links are explained in terms of community involvement of nodes, their roles and respective behavioral patterns. Our model allows for exploratory, descriptive and predictive tasks, including the analysis of communities, roles and their behavioral patterns, the interpretation of communities and roles as well as the prediction of missing links. Such tasks are enabled by posterior inference, for which we design a variational algorithm. An experimental assessment against state-of-the-art competitors reveals a higher accuracy of our approach on real-world benchmark data sets, both in community detection and link prediction. Role interpretation and behavioral role patterns are also demonstrated.