Mining the key predictors for event outbreaks in social networks
Chengqi Yi,Yuanyuan Bao,Yibo Xue
DOI: https://doi.org/10.1016/j.physa.2015.12.019
2016-01-01
Abstract:It will be beneficial to devise a method to predict a so-called event outbreak. Existing works mainly focus on exploring effective methods for improving the accuracy of predictions, while ignoring the underlying causes: What makes event go viral? What factors that significantly influence the prediction of an event outbreak in social networks? In this paper, we proposed a novel definition for an event outbreak, taking into account the structural changes to a network during the propagation of content. In addition, we investigated features that were sensitive to predicting an event outbreak. In order to investigate the universality of these features at different stages of an event, we split the entire lifecycle of an event into 20 equal segments according to the proportion of the propagation time. We extracted 44 features, including features related to content, users, structure, and time, from each segment of the event. Based on these features, we proposed a prediction method using supervised classification algorithms to predict event outbreaks. Experimental results indicate that, as time goes by, our method is highly accurate, with a precision rate ranging from 79% to 97% and a recall rate ranging from 74% to 97%. In addition, after applying a feature-selection algorithm, the top five selected features can considerably improve the accuracy of the prediction. Data-driven experimental results show that the entropy of the eigenvector centrality, the entropy of the PageRank, the standard deviation of the betweenness centrality, the proportion of re-shares without content, and the average path length are the key predictors for an event outbreak. Our findings are especially useful for further exploring the intrinsic characteristics of outbreak prediction.