Location and Language Independent Fake Rumor Detection Through Epidemiological and Structural Graph Analysis of Social Connections

Abstract

Detection and identification of misinformation and fake news is a complex problem that intersects several disciplines, ranging from sociology to computer science and mathematics. In this work, we focus on social media analyzing characteristics that are independent of the text language (language-independent) and social context (location-independent) and common to most social media, not only Twitter as mostly analyzed in the literature. Specifically, we analyze temporal and structural characteristics of information flow in the social networks and we evaluate the importance and effect of two different types of features in the detection process of fake rumors. Specifically, we extract epidemiological features exploiting epidemiological models for spreading false rumors; furthermore, we extract graph-based features from the graph structure of the information cascade of the social graph. Using these features, we evaluate them for fake rumor detection with 3 configurations: (i) using only epidemiological features, (ii) using only graph-based features, and (iii) using the combination of epidemiological and graph-based features. Evaluation is performed with a Gradient Boosting classifier on two benchmark fake rumor detection datasets. Our results demonstrate that epidemiological models fit rumor propagation well, while graph-based features lead to more effective classification of rumors; the combination of epidemiological and graph-based features leads to improved performance.