Machine Learning Interpretability to Detect Fake Accounts in Instagram

Abstract

This study is related to the detection of fake accounts on Instagram dataset that used by previous works. For this purpose, various Machine Learning algorithms have been used such as Bagging and Boosting to detect fake accounts on Instagram. Machine Learning now allows eight to learn directly from data rather than human knowledge, with an increased level of accuracy. To balance the two classes of data, we used the SMOTE algorithm which allows to obtain the same number of individuals for each class. We also incorporated methods for interpreting complex Machine Learning Models to understand the reasons for a model decision like SHAP values and LIME, we preferred to use SHAP values because it provides a local and global explanation of the model and also the values add up to the real estimation of the model, which LIME does not provide. Results show an overall accuracy of 96% for the XGBoost and Random Forest. In what follows, an online fake detecting system has been developed to detect malicious accounts on the Instagram.

A Fake Profile Detection Model Using Multistage Stacked Ensemble Classification

“…Elyusufi et al [9] employed DT and NB algorithms to detect fake profiles on social media and classified user profiles into genuine and fake. Sallah et al [16] proposed an ML architecture for detecting fake Instagram accounts using techniques like bagging and boosting, synthetic minority over-sampling technique (SMOTE), and SHapley Additive exPlanations (SHAP) values, with a combined accuracy of 96% using XGBoost and RF models. However, most of these studies lack accuracy and extensive evaluation and comparison with state-of-the-art classifiers.…”

“…Thus, in the proposed work on detecting fake profiles in OSN, seven base classifiers that are frequently used in the existing research on fake profile detection and found to be effective are chosen. These classifiers include DT [9], AdaBoost [16], KNN [10], LR [11], RF [8], NB [12], XGBoost [13], and SVM [14]. Each of these classifiers has its strengths in resolving classification problems, especially in fake profile detection.…”