An interpretable data-driven approach for modeling toxic users via feature extraction

Online Social Networks (OSNs) enable large-scale discussions but often suffer from toxic behaviors such as harassment and hate speech. While automated moderation helps manage toxicity, personalized approaches remain challenging due to fairness and transparency concerns. We introduce utoxic, a machine-learning framework that detects and analyzes toxic users based on linguistic, affective, and clustering-derived features. It performs binary and multi-class classification while incorporating explainability techniques for transparency. Evaluating utoxic on a Reddit dataset with over 8 million comments, we demonstrate its effectiveness in identifying toxic users and specific toxicity types. Our approach enhances automated moderation, offering interpretable insights for fairer and more adaptive interventions.