A Framework for Enhancing Social Media Misinformation Detection with Topical-Tactics

Abstract

Recent years have seen advancements in machine learning methods for the detection of misinformation on social media. Yet, these methods still often ignore or improperly incorporate key information on the topical-tactics used by misinformation agents. To what extent does this affect the (non)detection of misinformation? We investigate how supervised machine learning approaches can be enhanced to better detect misinformation on social media. Our aim in this regard is to enhance the abilities of academics and practitioners to understand, anticipate, and preempt the sources and impacts of misinformation on the web. To do so, this article leverages a large sample of verified Russian state-based misinformation tweets and non-misinformation tweets from Twitter. It first assesses standard supervised approaches for detecting Twitter-based misinformation both quantitatively (with respect to classification) and qualitatively (with respect to topical-tactics of Russian misinformation). It then presents a novel framework for integrating topical-tactics of misinformation into standard “bag of words”-oriented classification approaches in a manner that avoids data leakage and related measurement challenges. We find that doing so substantially improves the out-of-sample detection of Russian state-based misinformation tweets.