Comparison of Machine Learning Algorithms for Multilingual Health Hoax Classification on X and YouTube

Short health-related messages on social media can influence how users interpret diseases, vaccines, and treatment issues. Because this type of content often appears as brief posts or video titles, automatic classification is needed to help identify misleading claims. This research evaluates three machine learning algorithms, namely Naïve Bayes, Support Vector Machine, and Logistic Regression, for health-hoax classification on X and YouTube. The data were collected from January 2020 to January 2026 and consisted of Indonesian and English records. English texts were translated into Indonesian so that all records could be processed in one language pipeline. The processing stages included text cleaning, normalization, tokenization, stemming, stopword removal, IndoBERT-based pseudo-labeling, TF-IDF feature representation, Random Under Sampling, and parameter tuning using GridSearchCV. The final dataset contained 10,983 records, consisting of 7,157 non-hoax and 3,826 hoax records. The experimental results show that Support Vector Machine produced the highest weighted F1-score of 0.7467 in the 70:30 split scenario without balancing. Logistic Regression produced a close weighted F1-score of 0.7462 and obtained the highest AUC value of 0.8010. Random Under Sampling improved hoax recall, but it also reduced the overall weighted F1-score.