<![CDATA[Across Indonesian online platforms, fabricated news spreads faster than fact-checkers can confirm. Because much of the literature relies on resource-intensive deep models, one applied question stays unsettled: which lighter, more transparent classifier best detects Indonesian hoaxes? We assessed four algorithms, Random Forest (RF), Support Vector Machine (SVM), Naïve Bayes (NB), and Extreme Gradient Boosting (XGBoost), on 1,116 Indonesian articles from the MAFINDO/TurnBackHoax repository, using one shared preprocessing pipeline and two feature schemes, Bag-of-Words (BoW) and TF-IDF, under an 80:20 stratified split. On the held-out test set, Random Forest with BoW performed best at 98.66% accuracy and 98.59% macro F1-score, missing three of 224 cases, with XGBoost next at 98.21%. Under repeated five-fold cross-validation, however, XGBoost with BoW attained a significantly higher mean (98.36% versus 97.44% macro F1), so the two ensembles are best regarded as closely competitive rather than decisively ranked. BoW also beat TF-IDF for every classifier, most sharply for Naïve Bayes (84.70% versus 70.19% F1). Because the study uses lexical features alone on a single dataset, the findings indicate that classical ensembles can rival reported deep-learning figures at far lower cost, rather than forming a universal conclusion; broader validation across other sources and periods is still needed.]]>
Copyrights © 2026
