Efektivitas Logistic Regression dalam Analisis Sentimen Berbahasa Indonesia pada Komentar YouTube tentang Isu Ketenagakerjaan

Abstract

This study examines the development of a sentiment classification system for Indonesian-language YouTube comments addressing employment issues through the implementation of Logistic Regression algorithm. The research dataset comprises 2,755 comments extracted from a video themed "Job Seeker Stories," with 1,020 comments manually labeled into three sentiment categories: positive, neutral, and negative. The research methodology includes text preprocessing stages, feature transformation using TF-IDF, data splitting with stratified sampling, class imbalance handling through SMOTE, and hyperparameter optimization using GridSearchCV. Model evaluation yielded 44% accuracy with varying performance distribution across classes. The negative class demonstrated optimal performance with an F1-score of 0.55, while neutral and positive classes achieved scores of 0.34 and 0.29, respectively. Class distribution imbalance and implicit characteristics of positive comments became primary obstacles in the classification process. Research findings indicate that the combination of Logistic Regression, TF-IDF, and SMOTE has potential as a baseline method for sentiment analysis of Indonesian social media comments. Nevertheless, deep learning-based model development is necessary to improve accuracy and linguistic nuance interpretation capabilities. The analysis also identified negative sentiment dominance in public responses, reflecting societal concerns regarding the national employment situation.