<![CDATA[User reviews of online learning applications such as Ruangguru provide valuable information for evaluating service quality, user experience, and digital learning effectiveness. Although IndoBERT has demonstrated strong performance in Indonesian sentiment analysis, previous studies generally compared imbalance handling techniques using different datasets, model architectures, and experimental protocols, making the relative effectiveness of data-level and cost-sensitive learning approaches difficult to evaluate objectively under the same Transformer backbone. This study compares a Data-Level Approach using Latent-SMOTE with a Cost-Sensitive Learning Approach using Class-Weighted Loss on an identical IndoBERT architecture. The dataset consists of 3,767 Ruangguru user reviews collected from Google Play Store and processed through text preprocessing, IndoBERT tokenization, stratified train-validation-test splitting, and evaluation using Accuracy, Precision, Recall, Macro F1-score, confusion matrix, Cochran's Q Test, and McNemar Test. Experimental results show that the Baseline model achieved the highest Accuracy (90.05%), while the Cost-Sensitive Learning approach obtained the highest Macro F1-score (0.6275), outperforming both the Baseline (0.5658) and the Data-Level approach. These findings indicate that class-weighted optimization improves minority-class recognition without modifying the original training distribution, whereas Latent-SMOTE enhances minority representation but does not outperform Class-Weighted Loss. McNemar testing further confirms that the improvements over the Baseline are statistically significant. The main contribution of this work is an objective comparison of Data-Level and Cost-Sensitive Learning approaches using the same IndoBERT backbone, dataset, preprocessing pipeline, hyperparameters, and evaluation protocol. In addition, the study applies Latent-SMOTE in the latent feature space and complements performance evaluation with statistical significance testing, providing stronger empirical evidence for handling imbalanced Indonesian sentiment datasets]]>
Copyrights © 2026
