Creating a bot detector on social network

Abstract

The rapid growth of social media platforms has enabled new forms of digital interaction but has also given rise to malicious automated accounts, or bots, which threaten the integrity of online spaces. This thesis addresses the problem of detecting bots on Twitter using structured metadata features such as user activity, profile statistics, and account verification status. A comprehensive experimental study was conducted using various machine learning models (Random Forest, SVM, Naive Bayes, etc.) and deep learning architectures (RNN, LSTM, GRU). Each model was evaluated using accuracy, precision, recall, and F1-score metrics. The results show that Random Forest achieved the highest accuracy (86.89%) among all tested models, while LSTM led the deep learning group with 83.37% accuracy. These findings highlight the effectiveness of both traditional and deep learning approaches in classifying user accounts based solely on metadata. This research contributes to the field of social bot detection and lays the groundwork for future systems that enhance the security and authenticity of online platforms.