Cost-Sensitive Fake Profile Detection in Online Social Networks Using Random Forest Feature Selection and LightGBM

The proliferation of fake profiles on Online Social Networks (OSNs) presents serious risks to privacy, security, and trust. Traditional detection methods often struggle with large-scale data and fail to keep up with evolving tactics of malicious actors, highlighting the need for scalable, interpretable machine learning solutions. This study introduces a cost-sensitive and interpretable framework for identifying fake profiles by combining Random Forest (RF) feature selection with advanced gradient boosting models, specifically eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM). The framework was tested on the MIB Twitter dataset using a user-level split to prevent data leakage and ensure a realistic evaluation. Results show that LightGBM achieved the highest Cost-Sensitive Accuracy (CSA) of 0.96, surpassing XGBoost by 2.8% and RF by 4.6%, while training approximately 20% faster than XGBoost. These findings demonstrate that LightGBM strikes the best balance among predictive accuracy, cost sensitivity, and computational efficiency. By focusing on CSA as a key performance metric, this work highlights the importance of reducing false negatives in OSNs, where undetected fake accounts can cause more harm than false positives. Overall, the proposed framework offers a practical, scalable, and interpretable solution for real-time detection of fake profiles on online social networks. This study demonstrates that combining feature selection with cost-sensitive boosting effectively improves trust and security on large online social platforms.