Improving Sentiment Classification of Indonesian Skincare Reviews through Fine-Tuned IndoBERT and Data Augmentation

Abstract

This study analyzes sentiment in customer reviews of local skincare serum products on the Tokopedia e-commerce platform using a fine-tuned IndoBERT model enhanced with data augmentation techniques. A total of 5,000 reviews were collected from ten local skincare brands through web scraping and labeled according to star ratings into three sentiment classes: Positive, Neutral, and Negative. The dataset exhibited extreme class imbalance, with the Positive class representing 94.66% of all observations, creating substantial challenges for minority-class recognition. The data were divided through stratified sampling into 70% training, 15% validation, and 15% test sets to preserve class distributions. To mitigate imbalance, back-translation from Indonesian to English and back to Indonesian, together with synonym replacement, was applied exclusively to minority classes within the training set. The IndoBERT-base-p1 model was subsequently fine-tuned using focal loss combined with class weighting and compared against a baseline model trained without augmentation. Experimental results show that the proposed model achieved 94.40% accuracy, a Macro F1-score of 61.43%, and a Weighted F1-score of 95.14%. Although the baseline model obtained higher overall accuracy of 97.47%, it completely failed to identify the Neutral class, producing an F1-score of 0.00%. In contrast, the proposed approach increased the Neutral F1-score to 23.53% and improved the Macro F1-score by 2.30 percentage points, demonstrating more balanced performance across sentiment classes. The resulting model was deployed as SerumSense, a web-based application developed using Streamlit and SQLite, supporting both single-review and batch sentiment analysis. Black-box testing across 20 functional scenarios confirmed that all application features operated successfully as intended. These findings demonstrate that combining IndoBERT fine-tuning, targeted data augmentation, focal loss, and class weighting offers a practical approach for improving minority-class recognition in highly imbalanced Indonesian e-commerce review datasets.