Optimizing Sentiment Analysis on Imbalanced Hotel Review Data Using SMOTE and Ensemble Machine Learning Techniques

This research addresses the challenge of imbalanced sentiment classes in hotel review datasets obtained from Traveloka by integrating SMOTE (Synthetic Minority Oversampling Technique) with ensemble machine learning methods. The study aimed to enhance the classification of Positive, Negative, and Neutral sentiments in customer reviews. Data preprocessing techniques, including tokenization, stemming, and stopword removal, prepared the textual data for analysis. Various machine learning models—CART, KNN, Naive Bayes, and Random Forest—were evaluated individually and in ensemble configurations such as Bagging, Stacking, Soft Voting, and Hard Voting. The Stacking ensemble approach, utilizing Logistic Regression as a meta-classifier, demonstrated superior performance with an accuracy, precision, recall, and F1-score of 88%, outperforming Bagging (86%), Hard Voting (84%), and Soft Voting (81%). The findings highlight the effectiveness of SMOTE in balancing sentiment classes, particularly improving the classification of underrepresented Neutral and Negative categories. The novelty of this study lies in the comprehensive use of ensemble techniques combined with SMOTE, which significantly enhanced prediction stability and accuracy compared to previous approaches. These results provide valuable insights into leveraging advanced machine learning techniques for sentiment analysis, offering practical implications for improving customer experience and service quality in the hospitality industry.