SMOTE to Improve the Performance of Naïve Bayes and Random Forest in Sentiment Analysis of Digitalent applications

Abstract

Sentiment analysis is critical to understanding how an app, such as a digital training app like Digitalent, is viewed by users. User reviews available on app distribution platforms provide ample data for this analysis. However, in sentiment analysis, data imbalance is a common problem; positive reviews tend to outnumber negative and neutral reviews. This imbalance can impact machine learning models, which can lead to inaccurate predictions of the majority class. The purpose of this research is to solve this problem by using SMOTE (Synthetic Minority Selection Technique) technique in sentiment analysis of Digitalent app reviews and comparing the performance of two machine learning algorithms, Naive Bayes and Random Forest. The research data was collected from Indonesian user reviews from the Digitalent platform. Before being processed for analysis, the data went through pre-processing processes such as cleaning, tokenization, and normalization. SMOTE technique was applied to balance the number of reviews for each sentiment class. Furthermore, Naive Bayes and Random Forest algorithms are used to categorize the sentiment. The results of the SMOTE application research successfully increased the proportion of negative and neutral classes, so that the distribution of the dataset became balanced. The test results show that the accuracy of Naïve Bayes increased from 68.25% to 92.16%, while Random Forest increased from 68.25% to 92.16%.Keywords: K-Means Clustering, education level, clustering, village education, RapidMiner