<![CDATA[This study explores sentiment classification on Tokopedia user reviews using TF-IDF for feature extraction and SMOTE to handle class imbalance. From nearly one million raw reviews sourced from Kaggle ("E-Commerce Ratings and Reviews in Bahasa Indonesia"), a final set of 6,477 relevant entries was obtained after rigorous preprocessing, including case folding, noise removal (emojis, URLs, numbers), normalization to KBBI standards, tokenization, stopword removal, and stemming with Sastrawi. The dataset consisted of 5,213 positive and 1,264 negative reviews (80.4% positive). SMOTE balanced the classes to 10,426 reviews with a 1:1 ratio for training. Five traditional machine learning models were evaluated: Naive Bayes, Logistic Regression, Support Vector Machine (SVM), Decision Tree, and Random Forest. Assessments were based on accuracy, precision, recall, F1-score, ROC-AUC, and computational time, using an 80:20 stratified split and 5-fold cross-validation. Random Forest achieved the best overall performance (accuracy: 0.9163, F1-score: 0.9133, ROC-AUC: 0.9784), while tuned SVM (C=10, RBF kernel) attained the highest accuracy of 0.9473 and F1-score of 0.9321. Cross-validation on Naive Bayes showed consistent results with an average accuracy of 88.09%. Further analysis using Logistic Regression coefficients identified influential features: positive sentiment associated with words like "mantap", "mudah", and "sukses", while negative sentiment correlated with "kecewa", "parah", and "lemot". These insights provide practical value for Tokopedia's teams to enhance user experience, such as improving app speed and addressing complaints. The findings demonstrate the effectiveness and efficiency of traditional machine learning techniques for sentiment analysis in Bahasa Indonesia contexts.]]>
