INTERPRETABLE MACHINE LEARNING FOR DETECTION OF SPAMBOTS AND FAKE FOLLOWERS ON SOCIAL NETWORKS USING FEATURE BASED AND TEXT BASED

Abstract

Social networking sites like Twitter facilitate a wide interaction with humanity but are infiltrated more by robotics which imitating the human behavior propagate false information and control masses. To ensure the information integrity, Spambot detection is necessary, whereas much of the existing techniques rely on opaque, black box models, which restricts transparency and interpretability. In this paper, interpretable machine learning on spambot and false follower classification has been explored on the Cresci-15 and Cresci-17 datasets. Normalization, and tokenization, as well as the removal of extraneous material are part of preprocessing of text-based and feature-based data. SMOTE and SMOTEENN methods eliminate imbalance in the classification, and the RFE method reduces the feature dimension. They are the Decision Tree, the Random Forest, the SVM, the XGBoost, the AdaBoost, the Stacking Classifier and the Voting Classifier. These findings demonstrate that Stacking Classifier has the highest accuracy of 99.9 per cent in Cresci-15 and 99.5 per cent in Cresci-17. Explainable AI tools such as LIME and SHAP show the importance of features, which enhances transparency and decision-making with the model. Such a combination of feature selection, advanced resampling, and ensemble learning algorithms and interpretable methods of robust social network automatic recognition of accounts is demonstrated to be working.