Sentiment-based tourism review classification using support vector machine with recursive feature elimination and synthetic minority oversampling

User-generated content, particularly sentiment-rich reviews, is a critical data source for the tourism industry. However, analyzing this data presents significant computational challenges: (1) high- dimensional feature spaces from text (e.g., TF-IDF) which increase computational complexity, and (2) severe class imbalance, which skews classifier performance towards the majority class (e.g., positive reviews). While deep learning models have emerged, classical machine learning offers robust alternatives when properly optimized. This study proposes and evaluates a hybrid classification framework based on Support Vector Machine (SVM), chosen for its proven efficacy in high-dimensional text classification. To address the aforementioned challenges, the model integrates two key techniques: Recursive Feature Elimination (RFE) to select the most salient features and reduce computational load, and the Synthetic Minority Oversampling Technique (SMOTE) to create a balanced data distribution for training. Experimental results demonstrate the profound impact of this integrated approach. Data balancing with SMOTE significantly improved model performance, boosting baseline accuracy from 76% to 96% (without RFE) and 78% to 95% (with RFE). The final optimized model (SVM+RFE+SMOTE) achieved high- performance metrics of 96% accuracy, 96% precision, 97% recall, and a 97% F1-score. Furthermore, RFE successfully reduced computation time (e.g., from 70s to 40s on a 956-review set). This study concludes that the proposed SVM+RFE+SMOTE framework is a highly effective and efficient method for sentiment- based classification of real-world tourism reviews.