Feature identification for predicting community evolution in dynamic social networks

In parallel with the increasing popularity of commercial social-networking systems, the scales of such systems have grown notably, now with sizes ranging from hundreds of millions to more than a billion users. Besides being large, these systems also have a dynamic, temporal nature, with evolving structures. Thus, one of the main challenges is to understand and model the evolution of the meso-scale structures such as community structures within these networks. Most previous studies have concentrated on determining community events based on the community features extracted at different time points. However, both the huge volume of data and the dynamic structure of the networks hinder effective computation of these features. In this paper, we propose a novel framework that examines various structural features of the network and detects the most prominent subset of community features in order to predict the future direction of community evolution. Our approach is to extract the network structure and use it to determine the subset of community features that leads to accurate community event prediction. Unlike traditional approaches that harvest a large number of features at each time point, the proposed framework requires extraction of a minimal number of community features to effectively determine whether a community will remain stable or undergo certain events such as shrink, merge or split. Moreover, the extracted community features vary depending on the network structure, capturing network specific characteristics. Several experiments conducted on four publicly available datasets verified the effectiveness of the proposed framework.