Evaluasi Performa Algoritma Machine Learning untuk Klasifikasi Sentimen Ulasan Film IMDb

Abstract

IMDb with over 83 million active users generates millions of movie reviews annually. Manually analyzing such large textual data is impractical, necessitating automated approaches for binary sentiment classification. This study tested three traditional machine learning algorithms, namely Logistic Regression, Naive Bayes, and Random Forest, on 50,000 English-language movie reviews with balanced distribution between positive and negative sentiments. Feature extraction was performed using CountVectorizer and TF-IDF after preprocessing steps including HTML tag removal, contraction expansion, and text normalization. Among all combinations tested, Logistic Regression with TF-IDF achieved the best results with 87.29% accuracy, 87.32% F1-score, and 0.9481 AUC-ROC. Performance consistency was confirmed through 5-fold cross-validation with low coefficient of variation across all folds. Feature importance analysis revealed that words like "excellent," "perfect," and "amazing" indicate positive sentiment, while "worst," "awful," and "terrible" mark negative sentiment. Naive Bayes excelled in speed with 11.5 seconds training time but lower accuracy, while Logistic Regression offered the best balance between accuracy and efficiency at 17.2 seconds. Beyond accuracy, Logistic Regression provides interpretable coefficients explaining which words drive predictions. The findings demonstrate that properly tuned traditional algorithms remain viable for sentiment analysis in the entertainment industry