Learning pretopological spaces to extract ego-centered communities

Abstract

We present a pretopological based approach to extract ego-centered communities. Classical methods often consider only one structural feature of the network, whereas pretopology enables to do multi-criteria analysis. Our approach consists in learning a logical combination of network’s descriptors to define a pretopological space. Ego-centered communities are extracted by computing the elementary closure of each node. The quality of such communities is evaluated against the ground truth communities. We show the benefits of our method by comparing it to others on both real and synthetic networks.