March 24th14:30 , R2014 Digiteo Shannon (660) (see location):
Title : Community detection with modularity: a statistical physics approach.
Abstract :In community detection, the goal is to regroup nodes of an observed network into different groups (or communities) of nodes believed to be similar. With the rapidly growing number and size of real networks available, significant work has been made both to understand the fundamental limits in community detection, and to develop efficient algorithms. Many of them rely on the maximization of an indicator called modularity.
In this talk, I will first present the advantages and inconvenients of algorithms relying on modularity. I will then present a modularity-maximizing algorithm recently introduced by Zhang and Moore (2014). Called mod-bp, it solves some of these incovenients with methods inspired from statistical physics, introducing a temperature-like parameter T.
I will describe the behavior of this algorithm on synthetic and real networks and focuss on the influence of the temperature T. I will show the existence of different phases in which mod-bp gives qualitatively different results.
Finally, I will discuss how these findings can give new insights into community detection algorithms and into the structure of networks.
Contact: cyril.furtlehner at inria.fr