<![CDATA[The rapid growth of the skincare industry has sparked a surge in consumer reviews across digital platforms, including Sephora. These digital opinions hold valuable insights into public preferences regarding skincare products; however, the sheer volume of data poses a challenge for manual analysis. This study aims to classify the sentiment of skincare product reviews on Sephora by implementing the Support Vector Machine (SVM) algorithm. A dataset of 49,918 cleaned reviews was utilized following data preprocessing. Sentiment labels were automatically assigned based on ratings: scores of 4–5 were categorized as positive sentiment, while scores of 1–3 were categorized as negative sentiment. Text feature representation was constructed using Term Frequency-Inverse Document Frequency (TF-IDF), limited to the top 5,000 features. The data was subsequently split into training (80%) and testing (20%) sets using a stratified sampling approach. The SVM model employed a linear kernel with a parameter value of C = 1. Testing results demonstrated an accuracy of 90.49% and a macro F1-score of 84.70%. Specifically, the positive class achieved a precision of 92.19%, recall of 96.12%, and F1-score of 94.12%, whereas the negative class recorded a precision of 82.53%, recall of 69.22%, and F1-score of 75.29%. Comparative analysis revealed that SVM outperformed Naive Bayes and Logistic Regression in terms of accuracy and macro F1-score. However, the recall achieved for the negative class indicates a need for further optimization in detecting negative responses. This study confirms the effectiveness of integrating TF-IDF and linear SVM for skincare review sentiment classification, while noting that rating-based annotation has limitations in fully capturing the emotions expressed in the text.]]>
Copyrights © 2026
