Feature Selection Optimization for Sentiment Analysis of Tax Policy Using SMOTE and PSO

In public policy and governance, the issue of taxation often becomes a particular concern among the public. This research aims to understand public opinion on Direktorat Jenderal Pajak Indonesia’s performance. This research adopts a sentiment analysis approach, using a dataset of comments collected from the YouTube social media platform. One significant obstacle in this analysis is the imbalance of comment sentiment data, with either positive or negative sentiment dominating. Researchers applied SMOTE oversampling and Particle Swarm Optimization (PSO) techniques on feature selection to improve the quality of sentiment analysis models. SMOTE generates synthetic data from minority classes to ensure that the train data is balanced and the model does not contain bias due to data imbalance. This study proves that the methods are effective, especially in a 70% training data split scenario. At 70% split, the recall value increased from 0.47 to 0.52, a significant improvement in detecting minority sentiment that is often ignored in similar studies. The feature selection technique using PSO with an F1-score as the best criterion achieves substantial improvement on all evaluation metrics: accuracy reaches 0.93, recall 0.63, precision 0.70, and F1 score 0.66. This research proves the effectiveness of these methods in modeling various aspects of sentiment about taxation in Indonesia.