Unsupervised Fake News Detection

Abstract

Fake news has become more prevalent than ever, correlating with the rise of social media that allows every user to rapidly publish their views or hearsay. Today, fake news spans almost every realm of human activity, across diverse fields such as politics and healthcare. Most existing methods for fake news detection leverage supervised learning methods and expect a large labelled corpus of articles and social media user engagement information, which are often hard, time-consuming and costly to procure. In this paper, we consider the task of unsupervised fake news detection, which considers fake news detection in the absence of labelled historical data. We develop GTUT, a graph-based approach for the task which operates in three phases. Starting off with identifying a seed set of fake and legitimate articles exploiting high-level observations on inter-user behavior in fake news propagation, it progressively expands the labelling to all articles in the dataset. Our technique draws upon graph-based methods such as biclique identification, graph-based feature vector learning and label spreading. Through an extensive empirical evaluation over multiple real-world datasets, we establish the improved effectiveness of our method over state-of-the-art techniques for the task.