Analisis Active learning SVM berbasis Margin Sampling pada Sentimen YouTube MBG

Abstract

The Free Nutritious Meal Program (MBG) has generated extensive public discussion on YouTube, making sentiment analysis a valuable tool for understanding public perceptions. This study investigates the effectiveness of Support Vector Machine (SVM)-based Active Learning (AL) using a Margin Sampling strategy and a Human-in-the-Loop (HITL) framework, in which researchers acted as oracles following annotation guidelines for three sentiment classes: Negative, Neutral, and Positive. The dataset comprised 999 labeled Indonesian YouTube comments, split into 599 initial training samples, 200 oracle pool samples, and 200 fixed test samples. Texts were represented using TF-IDF features with unigrams and bigrams (max_features = 5,000). Three approaches were compared: AL-HITL, Simulated Active Learning, and Random Sampling. After 15 iterations involving 50 additional labeled samples, AL-HITL achieved the highest macro F1-score of 0.5389, outperforming Simulated AL (0.4950) and Random Sampling (0.4977). A skip mechanism with a 33.3% skip rate reduced negative learning, whereas the other approaches experienced performance declines from the baseline score of 0.5154. Positive sentiment, the minority class (20%), yielded the lowest F1-score (0.4286), while the majority negative class (40.7%) achieved the highest (0.6098). These findings provide preliminary empirical evidence on the potential and challenges of Margin Sampling-based Active Learning for Indonesian YouTube sentiment analysis.