Sentiment analysis in the Indonesian e-commerce sector faces significant challenges due to the informal nature of language and severe class imbalance, where neutral reviews are often underrepresented. This research proposes a hybrid framework combining the deep semantic capabilities of IndoBERT with the Synthetic Minority Over-sampling Technique (SMOTE) to improve classification fairness. Using a dataset of Tokopedia customer reviews, this study compares a baseline model against a balanced model using SMOTE on 768-dimensional IndoBERT features. The experimental results reveal that while the baseline model achieved a high overall accuracy of 83%, it suffered from an "accuracy paradox," exhibiting a dismal recall of only 0.07 for the neutral class. Upon implementing SMOTE, the neutral class recall surged to 0.29, marking a significant 314% improvement in minority class detection. Although overall accuracy slightly decreased to 81%, the Macro Average F1-Score increased from 0.61 to 0.65, proving that the model is more robust and objectively reliable across all sentiment polarities. This study demonstrates that sacrificing marginal accuracy for improved minority sensitivity is vital for providing accurate business intelligence in the digital marketplace. These findings provide a robust roadmap for developing more equitable automated sentiment analysis systems in Indonesia.