Implementasi Ensemble Voting Classifier untuk Analisis Sentimen Publik terhadap Kontroversi Pertandingan Indonesia vs Bahrain pada Platform Youtube

Abstrak

The qualifying match between the Indonesian National Team and Bahrain sparked diverse public responses that escalated into controversy, particularly in digital spaces such as YouTube comment sections. This phenomenon prompted research to analyze public sentiment using a machine learn-ing approach to understand trends and polarization of public opinion. The research data was ini-tially collected as many as 1,000 raw comments through the YouTube scraping process, which then went through a cleaning stage resulting in 984 valid comments. The research data was ob-tained through scraping YouTube comments, The valid data is then labeled using a combination of lexicon (au-to-suggest) and manual validation approaches into three categories, namely positive, negative and neutral. The preprocessing stage focused on normalizing non-standard language (slang) and handling negations to maintain contextual meaning. Next, feature extraction was per-formed using Feature Union, which combines word- and character-based TF-IDF, as well as nu-meric features such as text length and punctuation proportion. To address data imbalance, the SMOTE method was applied to improve minority class representation. The model used was an Ensemble Voting Classifier with a soft voting approach, which combines a calibrated Support Vector Machine, Logistic Regression, and Random Forest. Model optimization was performed us-ing GridSearchCV to obtain the best parameters. The evaluation results showed that the model performed well with an accuracy of 89.34%, a precision of 89.17%, a recall of 89.34%, and an F1-score of 89.18%. Furthermore, the application of SMOTE and negation handling has been shown to help reduce bias toward the majority class. The application of the SMOTE method has been shown to significantly improve model performance compared to the baseline model without oversampling, which only achieved an accuracy of 88.83%. Overall, this ensemble approach with multidimensional feature engineering is effective in producing an accurate sentiment analysis model for evaluating public response to sporting events.