Sentiment-Based Knowledge Discovery pada Aplikasi iPusnas Menggunakan Metode Machine Learning dan Deep Learning

Abstract

iPusnas is a digital library application developed by the National Library of the Republic of Indonesia since 2016, with over 1.5 million users. Despite its potential to improve literacy, the application has only received a rating of 2.0. This study conducted sentiment analysis on 7.596 reviews obatained through web scraping using the Google Play Scraper Library. The data then underwent preprocessing steps including case folding, data cleaning, tokenization, stopword removal, and stemming. Reviews were automatically labeled based on the rating score, where scores of 1-3 were categorized as negative, with 5.174 entries, and scores 4-5 as positive, with 2.422 entries. The dataset was split in an 80:20 ratio, with 80% for training, and 20% for testing. The machine learning models tested were SVM, Random Forest, CNN, LSTM, and RNN. The evaluation metrics included accuracy, precision, recall, F1-score, and confusion matrix. CNN and LSTM achieved the highest accuracy (82%), Random Forest and CNN achieved the highest precision (81%), RNN the highest recall (79%) and LSTM the highest F1-score (79%). McNemar test showed a significant difference between Random Forest and CNN, Random Forest and LSTM, and between RNN and LSTM, while CNN and LSTM, as well as CNN and RNN, showed no significant difference.iPusnas is a digital library application developed by the National Library of the Republic of Indonesia since 2016, with over 1.5 million users. Despite its potential to improve literacy, the application has only received a rating of 2.0. This study conducted sentiment analysis on 7.596 reviews obatained through web scraping using the Google Play Scraper Library. The data then underwent preprocessing steps including case folding, data cleaning, tokenization, stopword removal, and stemming. Reviews were automatically labeled based on the rating score, where scores of 1-3 were categorized as negative, with 5.174 entries, and scores 4-5 as positive, with 2.422 entries. The dataset was split in an 80:20 ratio, with 80% for training, and 20% for testing. The machine learning models tested were SVM, Random Forest, CNN, LSTM, and RNN. The evaluation metrics included accuracy, precision, recall, F1-score, and confusion matrix. CNN and LSTM achieved the highest accuracy (82%), Random Forest and CNN achieved the highest precision (81%), RNN the highest recall (79%) and LSTM the highest F1-score (79%). McNemar test showed a significant difference between Random Forest and CNN, Random Forest and LSTM, and between RNN and LSTM, while CNN and LSTM, as well as CNN and RNN, showed no significant difference.