HYBRID RF–XGBOOST STACKING ENSEMBLE FOR LOW-RESOURCE INDONESIAN TOXIC COMMENT CLASSIFICATION ON ROBLOX

Abstract

This study developed a hybrid stacking ensemble combining Random Forest and XGBoost for multi-class toxic comment detection in Indonesian gaming chat on Roblox. The dataset comprised 10,702 labeled comments across four toxicity categories: Violence, Harassment, Racist, and Neutral. Text preprocessing included comprehensive techniques such as case folding, tokenization, slang normalization, and stemming, while features were extracted using TF-IDF representation. The stacking ensemble achieved 91.4% accuracy and 91.3% macro F1-Score, significantly surpassing standalone baselines (Random Forest 87.2%, XGBoost 89.5%). McNemar’s test confirmed statistical significance of the improvement over XGBoost (p<0.001). Analysis revealed distinct linguistic patterns across toxicity categories. Importantly, the model operates at 2,500 predictions per second on CPU without GPU infrastructure, enabling practical deployment on resource-constrained gaming platforms. These findings demonstrate the effectiveness of hybrid ensemble learning for low-resource multilingual toxic content moderation.