<![CDATA[The rapid growth of social media in Indonesia has given rise to social phenomena such as Fear of Missing Out (FOMO). Expressions of FOMO on platforms like X (previously Twitter) often written informally, filled with abbreviations, slang, and emotional nuances, posing challenges for traditional Natural Language Processing (NLP) methods. This research aims to develop an optimized sentiment classification model for FOMO-related posts by fine-tuning the IndoBERT architecture and applying comprehensive data enhancement strategies. The study introduces three key innovations: (1) systematic text normalization to handle informal expressions, (2) a hybrid labeling framework combining automated model prediction, lexicon-based validation, and manual annotation to construct high-quality ground-truth data, and (3) hyperparameter tuning using both GridSearchCV for traditional machine learning models and Bayesian Optimization (Optuna) for deep learning models to maximize performance. The experimental results demonstrate that the optimized IndoBERT achieved superior performance with an Accuracy of 94.50%, F1-Score of 94.52%, and Macro AUC of 0.987. These results significantly surpass comparative models, including BiLSTM (Accuracy 86.60%), Support Vector Machine (88.06%), and Naive Bayes (80.73%). These results confirm that integrating hybrid labeling and fine-tuned IndoBERT significantly enhances sentiment classification performance. The findings contribute to developing reliable sentiment analysis systems for detecting social anxiety dynamics and computational social science research in Indonesian contexts.]]>
