SENTIMENT CLASSIFICATION MODEL BASED ON COMPARATIVE STUDIES USING MACHINE LEARNING TECHNOLOGY

Abstract

The development of social media has generated large amounts of text data, which is a valuable source for sentiment analysis. This study aims to conduct a comparative study of sentiment classification models on Indonesian-language YouTube comments, specifically comparing lexicon-based approaches, traditional machine learning models (Naive Bayes), and deep learning models (LSTM). Data was collected from YouTube videos themed around the youth generation and demographic bonuses, totaling 9,162 comments that underwent comprehensive text preprocessing. Model performance evaluation was conducted using accuracy, precision, recall, and F1-score metrics. The results show that the LSTM model outperforms Naive Bayes with an accuracy of 78.78% and an average F1-score of 0.79, compared to Naive Bayes, which only achieves an accuracy of 62.08% and an F1-score of 0.54. Although LSTM offers higher performance, the Naive Bayes model remains relevant due to its simplicity and efficiency. This study makes an important contribution to the selection of sentiment classification models for the Indonesian language and suggests the development of hybrid models and the use of contextual features for more optimal results. The LSTM model outperforms Naive Bayes with an accuracy of 82.15% (improved from 78.78% through enhanced regularization) and an average F1-score of 0.84. Comprehensive hyperparameter tuning via grid search and expanded manual annotation (40% of the dataset with κ=0.83) ensures robust model evaluation and reduces labeling bias. The study provides methodologically sound benchmarks for Indonesian sentiment analysis