<![CDATA[Sentiment analysis of e-commerce app reviews is essential to capture user perception and guide service improvements. However, review datasets are typically imbalanced—especially for the neutral class—making accuracy-only evaluation inadequate. This study proposes a hybrid approach that combines IndoBERT fine-tuning with a TF–IDF + logistic regression ensemble, augmented with probability calibration via temperature scaling, a dedicated neutral threshold rule, and a rating-based prior for low-confidence predictions. To avoid data leakage, the dataset is first split using stratified sampling into 72% training, 8% validation, and 20% testing; oversampling is applied only on the training split. Training uses label smoothing and early stopping (patience=2). The best validation configuration achieves macro-F1 of 0.8158 (T=0.941; α=0.70; t_neu=0.55; γ=0.10; τ=0.60). On the test set, the proposed model reaches 86.77% accuracy, 81.71% macro-F1, and 86.76% weighted-F1. An ablation study shows consistent gains from the TF–IDF+LR baseline to the full hybrid model, with the most notable improvement in the neutral class.]]>
Copyrights © 2026
