Detection of Body Shaming in Social Media: Traditional Machine Learning vs. Transformer-based Models

This paper explores the detection of body shaming content on social media by comparing traditional machine learning models with state-of-the-art transformer-based pre-trained language models. Two datasets were used: a 4K dataset from publicly available sources and an extended 6K dataset with data collected from TikTok and X. The traditional models assessed include Support Vector Machines, Logistic Regression, Naive Bayes, Random Forests, XGBoost, and AdaBoost, while the transformer-based models included BERT, RoBERTa, and XLNet. Performance was evaluated using precision, recall, macro-averaged F1-score, and Matthews Correlation Coefficient. Transformer-based models consistently outperformed traditional models, with BERT and RoBERTa achieving the best performance across datasets. The study highlights the superior ability of transformer models to capture complex linguistic patterns, offering better performance on larger and imbalanced datasets, while traditional models remain effective in resource-constrained settings or with smaller datasets.