<![CDATA[The digital transformation of social media platforms, particularly X (formerly Twitter), has been accompanied by an increase in digital fraud cases. Machine learning approaches are widely used for fraud detection; however, they face the challenge of class imbalance, in which fraudulent posts are substantially outnumbered by legitimate posts. This study aims to evaluate the effect of the Synthetic Minority Over-sampling Technique (SMOTE) on fraud detection performance and compare three classification algorithms: Naive Bayes, Support Vector Machine (SVM), and Logistic Regression. The data were obtained from the Kaggle dataset “ScamTags: Indonesian NER for Fraud Identification,” with features extracted using Term Frequency–Inverse Document Frequency (TF-IDF). To prevent data leakage, training and test data were strictly separated within the cross-validation procedure, ensuring that SMOTE was applied exclusively to the training folds and never to the test data. The experimental results show that, for this small and imbalanced dataset, SMOTE did not consistently improve model performance. Naive Bayes was the only model that effectively detected fraudulent posts, achieving the highest recall of 90.00% and an F-measure of 60.76% in the scenario without SMOTE. In contrast, SVM failed to detect the fraud class, yielding a recall of 0%, despite its seemingly high accuracy. These findings underscore the importance of leakage-free evaluation and highlight how accuracy can be misleading when dealing with imbalanced data. Consequently, minority-class recall and F-measure provide more informative measures for evaluating fraud detection systems. In practical terms, this study suggests that a simple model such as Naive Bayes can serve as a reliable initial option for screening Indonesian-language fraudulent posts, particularly when training data are limited. Nevertheless, these findings should be interpreted in light of the dataset's small size and single-source origin. Further validation using larger and more diverse datasets is required to establish the generalizability of the results.]]>
Copyrights © 2026
