Enhanced Bot Detection on TwiBot-20 Dataset

Bot detection is critical in safeguarding social networks against malicious activities such as propagating misinformation and shaping public opinion. Twitter, being extensively studied due to its accessibility and interactive nature, serves as an ideal platform for studying bot behaviors. In this study, we evaluate various network configurations using TwiBot-20 dataset to assess their efficacy in bot detection. Our investigation incorporates Graph Neural Networks (GNNs) models to leverage both network structure and textual content for enhanced detection accuracy. Experimental results demonstrate our combined model's superior performance, achieving high accuracy and robust metrics across the board. This research contributes to advancing bot detection methodologies, aiming to fortify social network integrity against emerging threats.

I. Introduction

Bot detection is crucial for today's social networks since bots have potential malicious use cases, such as imitating genuine users, spreading fake news, or even manipulative pro-paganda. Twitter is one of the most studied social networks among researchers due to its ease of data access and rich user interaction, which provides a sound ground for studying bot behaviors. Detecting these bots is vital for maintaining the platform's integrity. Effective bot detection relies on robust datasets, yet existing benchmarks often lack diversity and scale. To address this, we used the TwiBot-20 dataset, which includes 229,573 users, over 33 million tweets, exceeding 8.7 million user properties, and more than 455 thousand follow relationships [1]. This comprehensive dataset captures a wide spectrum of user types and detailed information. Also the amount of data provided in this dataset meets the quality requirements for bot detection as established by prior research [2], ensuring a reliable foundation for our experiments. In this study, we explore various network combinations using the TwiBot-20 dataset to evaluate their performance in bot detection. Our goal is to understand how different approaches fare in this context, contributing to the broader field of bot detection research.