A Deep Learning Method for Transformer-Based, Robust Bot Detection on Twitter

Abstract

Due to the large number of bots on Twitter, there has to be a way to reliably and accurately identify bots, both lawful and malevolent. These approaches worked, but they didn't solve the following problems: (1) the impossibility of obtaining ground truth real-world datasets due to the large datasets needed to train a model to detect bots; (2) the difficulty of learning representations of a diverse attributed network such as Twitter; and (3) the ongoing evolution of bot accounts to avoid automatic detection. In this study, we provide ADNET, a new framework for anomaly detection in networks ascribed to Twitter with little labeled data. Our proposed topology-based active learning framework, which trains the model using a deep autoencoder and outperforms prior techniques in handling huge graphs, is an attempt to remedy the shortcomings of earlier approaches. While reducing the annotation cost in Twitter attribution networks, our experimental findings show that the suggested strategy outperforms state-of-the-art approaches in identifying anomalous bot accounts.