Identifying Unconvincing User on Social Media with Limited Features

Unconvincing users on social media are responsible for many malicious online activities. Therefore, developing classifiers to identify them is crucial. Due to the trade-off between the crawling costs of features and sample size, it remains challenging to distinguish unconvincing users from normal human users using limited features with low API call cost. To address the challenge, we present a case study where we develop machine learning-based classifiers with limited features to specifically identify a predefined category of unconvincing users: fake Twitter followers. Four models, namely Random Forest (RF), XGBoost, LightGBM, and Support Vector Machine (SVM) are employed and trained using a tabular dataset consisting of only 19 low-cost features. Experimental results show that the LightGBM classifier achieves the highest accuracy at 98.7%. Furthermore, feature importance analysis is carried out on LightGBM, and both SHAP analysis and Information Gain (IG) results indicate that the ratio between friends and followers is the most important feature.