Comparing XGBoost and LightGBM for Optimizing Health Content Categories

Abstract

Indonesian social media platforms host a rapidly expanding flow of health-related information, much of it unverified and fragmented across major disease topics such as diabetes, heart disease, and hypertension. This study develops a supervised multi-class text-classification pipeline that integrates IndoBERT embeddings with LightGBM and XGBoost to identify the most effective model for disease-based health content categorization. Preprocessing includes data anonymization, normalization, tokenization, and contextual embedding extraction using pretrained IndoBERT; evaluation employs five-fold stratified cross-validation to maintain class balance. Performance is measured through accuracy, precision, recall, and macro F1-score supported by confusion matrices. Results show that IndoBERT × LightGBM achieves the highest accuracy (0.8526) and balanced macro F1 (0.85), outperforming IndoBERT × XGBoost (accuracy 0.8325, F1 0.81). Diagnostic results indicate that LightGBM’s leaf-wise boosting structure improves generalization on short, noisy Indonesian texts. Feature-importance analysis highlights contextual terms such as “blood sugar,” “heart,” and “blood pressure” as key linguistic indicators contributing to model predictions. The workflow provides an explainable, scalable baseline for health text monitoring, misinformation detection, and public-health analytics in low-resource language settings.

GS Cited Analysis