Improving Random Forest Performance for Sentiment Analysis on Unbalanced Data Using SMOTE and BoW Integration: PLN Mobile Application Case Study

Abstract

Purpose: This research aims to improve the accuracy of sentiment analysis on PLN Mobile app reviews by overcoming the challenge of data imbalance. This goal is important to provide a better understanding of user opinions and support PT PLN (Persero) in improving mobile application services.

Methods: This research uses the Random Forest algorithm combined with Synthetic Minority Over-sampling Technique (SMOTE) to handle imbalanced data. Data is collected through web scraping reviews from the Google Play Store, followed by preprocessing processes such as data cleaning, stopword removal, tokenization, and stemming. Feature extraction is performed using the Bag of Words (BoW) method, and the data is tested with four sharing schemes.

Result: The results showed that the 90%-10% sharing scheme gave the best performance with an accuracy of 81% and an average precision and recall of 0.79. This finding confirms that the larger the proportion of training data, the better the model performs sentiment classification.

Novelty: This research's novelty lies in combining SMOTE with BoW and Random Forest to overcome data imbalance. This approach is a significant reference for future sentiment analysis research. It provides practical insights that PT PLN (Persero) can use to improve the quality of its application services.