Comparison of IndoBERT and Naive Bayes Performance on Sentiment Analysis of Indonesian-Language E-Commerce Reviews

Abstract

The growth of e-commerce platforms in Indonesia has generated a large number of user reviews containing important information about service quality and user satisfaction. Manual analysis of these reviews is inefficient, necessitating a Natural Language Processing (NLP)-based approach. This study aims to compare the performance of Naive Bayes and IndoBERT in sentiment classification of Indonesian-language e-commerce reviews. The dataset consists of 26,103 reviews of Tokopedia, Shopee, and Lazada apps obtained from the Google Play Store. The research stages include sentiment labeling based on ratings, preprocessing, Term Frequency–Inverse Document Frequency (TF-IDF) weighting on Naive Bayes, fine-tuning IndoBERT, and evaluation using a confusion matrix with accuracy, precision, recall, and F1-score metrics. The results show that Naive Bayes achieved an accuracy of 80.53%, while IndoBERT achieved 86.48% with a Macro F1-Score of 0.6515 and a Weighted F1-Score of 0.8723. These results show that IndoBERT is more effective in understanding the context of Indonesian text, resulting in better sentiment classification performance.