Predicting Virality on Networks Using Local Graphlet Frequency Distribution

Abstract

The task of predicting virality has far-reaching consequences, from the world of advertising to more recent attempts to reduce the spread of fake news. Previous work has shown that graphlet distribution is an effective feature for predicting virality. Here, we investigate the use of aggregated edge-centric local graphlets around source nodes as features for virality prediction. These prediction features are used to predict expected virality for both a time-independent Hawkes model and an independent cascade model of virality. In the Hawkes model, we use linear regression to predict the number of Hawkes events and node ranking, while in the independent cascade model we use logistic regression to predict whether a k-size cascade will multiply by a factor X in size. Our study indicates that local graphlet frequency distribution can effectively capture the variances of the viral processes simulated by Hawkes process and independent-cascade process. Furthermore, we identify a group of local graphlets which might be significant in the viral processes. We compare the effectiveness of our methods with eigenvector centrality-based node choice.