Collective Classification for Social Media Credibility Estimation

Abstract

We introduce a novel extension of the iterative classification algorithm to heterogeneous graphs and apply it to estimate credibility in social media. Given a heterogeneous graph of events, users, and websites derived from social media posts, and given prior knowledge of the credibility of a subset of graph nodes, the approach iteratively converges to a set of classifiers that estimate credibility of the remaining nodes. To measure the performance of this approach, we train on a set of manually labeled events extracted from a corpus of Twitter data and calculate the resulting receiver operating characteristic (ROC) curves. We show that collective classification outperforms independent classification approaches, implying that graph dependencies are crucial to estimating credibility in social media.