<![CDATA[The development of online transportation applications such as Maxim has increased the need for sentiment analysis to understand user opinions from reviews on the Google Play Store. The main challenges in this analysis are language diversity, variations in writing style, and data imbalance, which affect model accuracy. This study aims to evaluate the performance of the Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Random Forest (RF) algorithms, as well as ensemble approaches through the Voting Classifier and Combined Classifier, in sentiment analysis of Maxim app reviews. The dataset consists of 2,851 Indonesian-language reviews collected through web scraping from the Google Play Store in 2025. Sentiment labels were automatically determined based on user ratings, where ratings of 4–5 were categorized as positive and ratings below 4 as negative, with an initial distribution of 2,295 positive and 556 negative reviews before balancing using SMOTE–Tomek Links. Preprocessing steps included case folding, tokenization, stopword removal, and stemming using Sastrawi, while feature weighting was performed with unigram TF-IDF. The Combined Classifier merged the probability scores from the SVM, KNN, and RF models to produce the final prediction. Evaluation was conducted using 5-Fold Cross Validation with accuracy, precision, recall, F1-score, and ROC-AUC as evaluation metrics. The results show that RF and the Combined Classifier achieved the best performance with 85% accuracy, 87% precision, 85% recall, 86% F1-score, and 0.91 ROC-AUC, while SVM and the Voting Classifier ranked in the middle and KNN ranked the lowest. These findings confirm that ensemble learning, particularly the Combined Classifier, effectively improves the accuracy and stability of review classification compared to individual methods.]]>
Copyrights © 2025
