Machine Learning Identification of High-Conflict Couples at Risk for Intimate Partner Violence

Abstract

This study aimed to develop and interpret machine learning models capable of identifying high-conflict couples at elevated risk for intimate partner violence by integrating multidimensional dyadic, psychological, and relational data. A cross-sectional predictive design was employed with a sample of 368 high-conflict heterosexual couples recruited from counseling and community support settings in Italy. Partners independently completed validated self-report measures assessing conflict dynamics, attachment orientations, emotional regulation, perceived stress, jealousy, relationship satisfaction, and intimate partner violence risk, alongside demographic information. Dyadic data were preprocessed and structured to preserve partner-level and couple-level information. Multiple supervised machine learning algorithms, including regularized logistic regression, support vector machines, random forest, and gradient boosting, were trained and evaluated using stratified cross-validation. Model interpretability was examined using explainable artificial intelligence techniques based on feature attribution. Ensemble-based models outperformed linear and kernel-based approaches, with the gradient boosting model demonstrating the highest predictive accuracy and discrimination (accuracy = 0.88; AUC = 0.94). Sensitivity to high-risk classifications was robust across ensemble models, indicating effective identification of couples at elevated risk. Feature importance analyses revealed that conflict escalation, anger dysregulation, attachment anxiety, perceived stress, and jealousy intensity were the strongest contributors to risk classification, while lower relationship satisfaction showed a smaller but meaningful effect. The results indicated that nonlinear interactions among relational and emotional variables substantially enhanced predictive performance. The findings demonstrate that explainable machine learning models can reliably identify high-conflict couples at risk for intimate partner violence by capturing complex dyadic interaction patterns. Integrating such models into preventive and clinical contexts may support earlier detection, targeted intervention, and ethically informed decision-making, complementing traditional assessment approaches.