<![CDATA[The Indonesia–United States Trade Agreement has sparked much public debate on YouTube, but due to the large volume of data and the informal nature of the text, manual analysis is ineffective. Through a comparative study of machine learning (Naïve Bayes and Support Vector Machine) and deep learning (Long Short-Term Memory and IndoBert), this research aims to identify the polarity of public opinion and evaluate the best computational model. This study contributes a comprehensive empirical comparison of the four algorithms within an identical experimental pipeline on an issue that has not been previously explored, thereby offering methodological insights for future sentiment analysis research on similarly informal, domain-specific social media text. The dataset consists of 2,284 comments labeled using the InSet Lexicon. The analysis results show that the neutral class dominates the sentiment distribution (52,0%), followed by the positive class (20,2%) and the negative class (27,8%). Performance evaluation reveals that IndoBERT significantly outperforms other models with an accuracy of 87.71% and a Macro F1-Score of 0.87. Support Vector Machine ranked second (79.00%), followed by Naïve Bayes (61.71%). In contrast, Long Short-Term Memory failed completely (accuracy of 52.00%) due to the “majority class collapse” phenomenon in small-scale datasets. This study concludes that the Transformer architecture (IndoBert) is the most robust approach for classifying sentiment in informal Indonesian-language text containing specific geopolitical and economic terms.]]>
