<![CDATA[This study investigates sentiment classification on imbalanced Indonesian user-review data from the MyIM3 mobile application, focusing on the effectiveness of hierarchical data augmentation for improving minority-class detection. A total of 10,000 reviews were collected, of which 6,968 were retained after preprocessing and automatically labeled using the InSet Indonesian sentiment lexicon. The labeled dataset was divided into training and testing sets (70:30) and used to fine-tune an IndoBERT-based sentiment classifier. Baseline results show that IndoBERT achieves strong overall performance (accuracy = 0.93, macro-F1 = 0.89, ROC–AUC = 0.969), however, detection of the minority positive class remains limited due to class imbalance. To address this issue, three augmentation strategies were evaluated: synonym replacement, random deletion, and hierarchical augmentation. Among these, hierarchical augmentation provides the most consistent improvement, increasing the positive-class F1-score from 0.81 to 0.83 (+0.02 absolute) while maintaining stable performance on the majority class. The findings indicate that hierarchical augmentation improves contextual sentiment learning by preserving clause-level structure while introducing controlled lexical variation, which is important for Indonesian evaluative expressions. This study contributes evidence to low-resource and imbalanced text classification research by demonstrating that structurally guided augmentation enhances minority sentiment recognition without degrading majority-class accuracy in Indonesian review datasets.]]>
