<![CDATA[Political misinformation spreads rapidly through digital media, making early detection increasingly important. This study evaluates headline-based Indonesian political misinformation classification using Naive Bayes (NB), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and IndoBERT. Headlines were selected because they represent the first information encountered by readers and enable rapid screening, although they may not fully represent the article content. The dataset consists of 5,132 Indonesian political news headlines collected from TurnBackHoax and DetikNews Politics. Leakage-aware preprocessing was applied to reduce explicit source-related cues before model training. Experimental results show that IndoBERT achieved the best performance with an accuracy of 90.25% and a Macro F1-score of 90.24%, outperforming the other evaluated models. Additional analyses, including McNemar statistical significance testing, computational cost comparison, and error analysis, showed that although IndoBERT achieved the highest predictive performance, its improvement over the other evaluated models was not statistically significant. Error analysis further revealed that several misclassification cases involved stylistic overlaps between misinformation and factual headlines, suggesting the possibility of residual source-style dependency and shortcut learning. Therefore, the proposed models should be interpreted as learning linguistic patterns associated with the constructed headline dataset rather than performing direct factual verification of news claims.]]>
