Speaker
Alcides Viamontes Esquivel
(IceLab, Umeå University)
Description
Infomap clusters networks using the correspondence between
compression, regularity detection and learning expressed in
the minimum description length principle. Here we extend
that approach to find partitions with overlaps, considering
flow on the nodes near the boundary of modules.
Specifically, we analyze how affects the modular structure
the return proportion of flow on those nodes, versus the
proportion going to different modules. We show that the
return proportion characteristic can be better captured --
in terms of average bit-length per step -- by allowing nodes
to belong to several modules, effectively making the modules
to overlap. This work introduces both the updated framework
and a fast greedy algorithm that finds the module
overlappings. Also, we present the outcomes of our new
method when processing several real world networks and in
the context of a benchmark procedure. For the later one, we
devised a way of calculating the mutual information between
two partitions that deals with overlaps consistently.
