Robust Sentiment Classification Using XGBoost with Imbalanced Learning for Social Media Analytics

Abstract

The rapid growth of social media platforms has generated massive volumes of user-generated content that provide valuable insights into public opinions and digital interactions. However, sentiment classification on social media data remains challenging due to noisy textual information, high-dimensional features, and severe class imbalance that reduces the ability of machine learning models to recognize minority sentiment categories. This study aims to develop a robust sentiment classification framework using XGBoost with imbalanced learning techniques to improve classification performance on social media analytics tasks. The research focuses on integrating feature extraction, dimensionality reduction, data balancing, and ensemble learning strategies to enhance sentiment prediction accuracy. A quantitative experimental approach was conducted using Coretax-related social media datasets by applying Term Frequency–Inverse Document Frequency (TF-IDF) for text representation, Truncated Singular Value Decomposition (SVD) for feature optimization, BorderlineSMOTE for minority-class augmentation, and XGBoost for sentiment classification. The results demonstrate that the proposed framework improves macro F1-score performance compared with baseline models, increasing from 0.4560 using Random Forest to 0.4776 with the proposed approach, and further improving to 0.4874 through emoji normalization. The findings confirm that combining feature engineering, imbalance-aware learning, and optimized preprocessing can enhance minority sentiment detection and support reliable social media intelligence. This study contributes to robust artificial intelligence and natural language processing research by addressing challenges in imbalanced sentiment classification. Future research should explore transformer-based models, explainable AI techniques, multilingual sentiment analysis, and real-time social media analytics frameworks.