Speaker
Mr
Johan Dahlin
(Swedish Defense Research Agency)
Description
During recent years many different methods to detect
community structures in complex networks have been
developed. Despite significant efforts from many different
scientists from different fields, no completely satisfactory
method to detect communities has been developed. In this
talk, we present a method to merge the results from several
different community detections run into a single final
estimated community structure. Each run can either use a
different method or else an ensemble of results from the
same algorithm an be merged. We propose three different
methods for the merge. Two of these are related to ensemble
clustering methods used in standard data clustering problems.
We apply these three methods on some different problems to
demonstrate their usefulness. Some existing methods are
stochastic in nature and generate different results for each
run. Using the merging methods, one can use the collective
information from the entire ensemble of possible community
structures to find the most likely structure. This is shown
to improve the performance of some stochastic algorithms.
Another problem is related to the problem of uncertain
Social Networks. In these networks, edges are not known with
certainty to exist. Instead, a probability (or probability
interval) is given for their existence. Using methods from
statistical simulation, one can create an ensemble of
networks that are consistent with the uncertain network.
Applying existing methods for detecting community structures
on each network from the ensemble creates a large number of
different candidate community structures (one for each
realization of the uncertain network). Applying the merging
methods presented in this talk, one can merge (fuse) these
different candidate structures into one, thus finding the
community structure of the uncertain network.
