Community detection and tracking on networks from a data fusion perspective

TL;DR

This paper introduces a probabilistic approach to community detection and tracking in networks, emphasizing uncertainty retention and Bayesian filtering to improve accuracy and applicability over traditional methods.

Contribution

It proposes a novel data fusion perspective that estimates community membership probabilities and develops Bayesian filters for dynamic community tracking.

Findings

The method accurately estimates community probabilities on the LFR testbed.

It is computationally efficient on standard network datasets.

The approach offers versatile applications complementing traditional hard-call methods.

Abstract

Community structure in networks has been investigated from many viewpoints, usually with the same end result: a community detection algorithm of some kind. Recent research offers methods for combining the results of such algorithms into timelines of community evolution. This paper investigates community detection and tracking from the data fusion perspective. We avoid the kind of hard calls made by traditional community detection algorithms in favor of retaining as much uncertainty information as possible. This results in a method for directly estimating the probabilities that pairs of nodes are in the same community. We demonstrate that this method is accurate using the LFR testbed, that it is fast on a number of standard network datasets, and that it is has a variety of uses that complement those of standard, hard-call methods. Retaining uncertainty information allows us to develop a Bayesian filter for tracking communities. We derive equations for the full filter, and marginalize it to produce a potentially practical version. Finally, we discuss closures for the marginalized filter and the work that remains to develop this into a principled, efficient method for tracking time-evolving communities on time-evolving networks.