Stacking of machine learning classifiers for bot detection using account level data

Abstract

Social media is a platform for individuals to connect, share, and create information. Social bots produce automated content and interact with humans; in the process, they learn and mimic humans’ behaviour. This research study addresses the challenge of identifying social media bots (SMB) that can rapidly disseminate information or misinformation on platforms like Twitter. It contributes to the field by reviewing literature to define bot behaviours and exploring advanced machine learning classifiers for effective bot detection using account-level data. The study employed Spearman's rank correlation coefficient to select relevant features for SMB classification, then trained six different machine learning models: decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), and k-nearest neighbour (KNN). To further improve accuracy, a classifier stacking technique was applied. Key findings revealed that while individual classifiers performed variably, with RF leading at 89% accuracy, the stacked classifier approach outperformed all single-classifier methods with an impressive 90% accuracy rate. The results underscore the potential of combining multiple classifiers to enhance the precision of social media bot detection efforts.