A Comparison of Computational Approaches for Sentiment Analysis on Public Opinions About Education in Indonesia

This study investigates sentiment analysis of education-related tweets in Indonesia. The dataset comprises 484 manually labelled tweets, exhibiting a significant class imbalance, with 65.5% negative, 29.1% neutral, and only 5.4% positive sentiments. Term Frequency-Inverse Document Frequency (TF-IDF) and Word2Vec were evaluated in combination with Support Vector Machine (SVM), Naive Bayes, and Long-Short Term Memory (LSTM) models. Experimental results demonstrated that the optimized SVM with linear kernel and TF-IDF vectorization achieved the highest accuracy of 80% and achieved an MCC score of 0.56, while Naive Bayes reached 75% after hyperparameter tuning, LSTM models showed consistent but lower performance at 66% accuracy across both vectorization methods. The study reveals the effectiveness of traditional machine learning approaches, particularly SVM with TF-IDF vectorization, for sentiment analysis of Indonesian educational tweets, while also highlighting challenges posed by class imbalance in the dataset. These findings contribute to advancing sentiment analysis approaches for Indonesian language social media data, especially in the educational domain.