Enhanced Twitter Bot Detection via Static and Temporal Feature Integration

Abstract

This study presents a robust supervised learning framework tailored for the precise identification of Twitter bots. The central contribution lies in the integration of both static and temporal features to establish a comprehensive representation of user behavior. Using the benchmark Cresci-2017 dataset, the framework constructs a rich feature space encompassing profile-level metadata, detailed posting frequency statistics, and fine-grained temporal activity features, enabling effective discrimination between sophisticated bots and genuine human users. To fully exploit this diverse information, an advanced stacking ensemble is employed. This architecture strategically combines the predictive strengths of multiple base learners Random Forest, LightGBM, XGBoost, and Support Vector Machine while a Logistic Regression meta-learner is trained to optimally integrate their outputs. The resulting model delivers state-of-the-art performance, achieving 98.71% accuracy, 99.26% precision, 99.04% recall, and a 99.15% F1-score, substantially outperforming individual classifiers. These findings underscore the value of unifying heterogeneous feature types with stacked generalization, highlighting that strategically diverse feature engineering coupled with advanced ensemble methods is essential for building resilient defenses against the rapidly evolving landscape of automated bot behavior.