Sentiment Analysis of Maxim App User Reviews in Indonesia Using Machine Learning Model Performance Comparison

Abstract

User reviews can vary widely in language and writing style, which can make accurate sentiment modeling difficult. Selecting the right machine learning model and comparing performance between models can be challenging, given that each model has its own strengths and weaknesses. The method used involved data collection by scraping 5000 reviews from the Google Play Store, followed by data pre-processing including data cleaning, tokenization, stemming, and feature engineering using TF-IDF. The data was divided into training (70%) and testing (30%) sets, with the SMOTE oversampling technique applied to address class imbalance. Three machine learning models were used: Random Forest, Support Vector Machine (SVM), and Naive Bayes. The results showed that the majority of reviews were positive, with a high average app rating. Word cloud analysis revealed that “service”, “driver”, “price”, and “time” were the most frequently discussed aspects in the reviews. In terms of model performance, SVM performed the best with an accuracy of 91.3%, followed by Random Forest (89%) and Naive Bayes (78%). Maxim was generally well received by users in Indonesia, with the majority of reviews being positive. The SVM model proved to be the most effective in classifying review sentiment, outperforming other models in accuracy and precision.