A Voting Ensemble Model and Explainability Framework for Fake Account Detection in Social Media

In today’s digital era, people are connected through various online social media platforms. While social media can serve many positive purposes, it can also be manipulated to distract users by spreading misinformation. This misinformation can originate from genuine accounts or fake accounts intentionally created to mislead users. Consequently, it is essential to detect fake accounts on social media platforms. However, the existing fake profile detection models often lack explanability making it difficult for users to make decisions. To tackle these challenges we proposed a methodology with Instafake, a publicly accessible dataset. We utilize a voting classifier technique that integrates the outputs of various models: Random Forest, Extra Trees, GaussianNB, and MLPClassifier. This combination leverages the strengths of each model to classify samples as either fake or real profiles accurately. To address the imbalanced nature of our dataset, we implement the SMOTE oversampling method to generate synthetic samples, which enhances the model’s generalization capability. Additionally, we apply feature engineering techniques, including follower-following ratios, polynomial feature transformations, and mutual information-based feature selection, to improve the model’s predictive performance. Our proposed methodology achieves an impressive F1-score of 98%, significantly surpassing baseline models. Furthermore, we incorporate advanced techniques for model explainability, such as permutation importance and SHAP explanations, to promote transparency and facilitate broader adoption of our model. Index Terms—fake profiles, fake account detection, LIME, SHAP, Instagram, Online social networks