Behavioral Forensics in Social Networks: Identifying Misinformation, Disinformation and Refutation Spreaders Using Machine Learning

TL;DR

This paper introduces a behavioral forensics approach using machine learning and deep learning-based graph embeddings to identify malicious actors spreading misinformation, disinformation, or refutation in social networks, based on their behavioral actions.

Contribution

It proposes a novel labeling mechanism for classifying users into five behavioral categories and leverages network features for improved detection of disinformation spreaders.

Findings

Achieved 77.45% precision and 75.80% recall in identifying malicious actors.

Demonstrated effectiveness of deep learning-based graph embeddings in behavioral analysis.

Extended binary classification to a five-category behavioral model.

Abstract

With the ever-increasing spread of misinformation on online social networks, it has become very important to identify the spreaders of misinformation (unintentional), disinformation (intentional), and misinformation refutation. It can help in educating the first, stopping the second, and soliciting the help of the third category, respectively, in the overall effort to counter misinformation spread. Existing research to identify spreaders is limited to binary classification (true vs false information spreaders). However, people's intention (whether naive or malicious) behind sharing misinformation can only be understood after observing their behavior after exposure to both the misinformation and its refutation which the existing literature lacks to consider. In this paper, we propose a labeling mechanism to label people as one of the five defined categories based on the behavioral actions they exhibit when exposed to misinformation and its refutation. However, everyone does not show behavioral actions but is part of a network. Therefore, we use their network features, extracted through deep learning-based graph embedding models, to train a machine learning model for the prediction of the classes. We name our approach behavioral forensics since it is an evidence-based investigation of suspicious behavior which is spreading misinformation and disinformation in our case. After evaluating our proposed model on a real-world Twitter dataset, we achieved 77.45% precision and 75.80% recall in detecting the malicious actors, who shared the misinformation even after receiving its refutation. Such behavior shows intention, and hence these actors can rightfully be called agents of disinformation spread.