<![CDATA[The rapid growth of the local skincare industry in Indonesia has intensified consumer interactions on Instagram and TikTok, generating a substantial volume of comments that are inefficient to analyze manually. This study aims to implement the IndoBERT model to classify sentiment in beauty product comments using the CRISP-DM framework. A total of 5,062 comments from five local brands (Somethinc, MS Glow Beauty, Avoskin, N.Pure, and SkinGame) were retrieved via web scraping and automatically labeled into positive and negative classes using a lexicon-based approach. The evaluation results demonstrate that the fine-tuned IndoBERT model achieved an accuracy of 87% under the 80:20 data-splitting scenario and 86% under the 75:25 scenario. In the 75:25 scenario, the model attained a precision of 84%, recall of 94%, and F1-score of 89%, while remaining free from overfitting as shown by the loss curve. In conclusion, the IndoBERT architecture demonstrates high reliability in classifying consumer sentiments despite challenges from slang, dermatological terms, and code-mixing.]]>
Copyrights © 2026
