Machine Learning-Based Classification of High-Conflict Couples Using Psychological, Relational, and Behavioral Interaction Features

Abstract

The present study aimed to develop and evaluate machine learning models for the classification of high-conflict couples using psychological characteristics, relational functioning indicators, and behavioral interaction features while identifying the most influential predictors of relationship conflict status. This cross-sectional predictive study was conducted among 624 couples (N = 1,248 individuals) recruited from multiple urban and suburban regions of Malaysia. Participants completed a comprehensive assessment battery measuring depression, anxiety, stress, attachment insecurity, emotional intimacy, dyadic adjustment, and conflict behaviors. In addition, couples participated in structured conflict discussion tasks that were coded for behavioral interaction features, including criticism, defensiveness, contempt, stonewalling, positive affect, and conflict resolution attempts. Following data preprocessing procedures, including normalization, missing-value imputation, and feature engineering, the dataset was divided into training and testing subsets using stratified sampling. Several supervised machine learning algorithms, including Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, Artificial Neural Network, and Extreme Gradient Boosting (XGBoost), were trained and compared. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). SHapley Additive exPlanations (SHAP) analyses were conducted to determine feature importance and model interpretability. All machine learning models demonstrated satisfactory predictive performance; however, ensemble learning algorithms significantly outperformed traditional approaches. XGBoost emerged as the best-performing model, achieving an accuracy of 94.1%, precision of 93.7%, recall of 93.3%, F1-score of 93.5%, and an AUC-ROC value of 0.978. Feature importance analyses revealed that dyadic adjustment, emotional intimacy, criticism frequency, psychological aggression, attachment anxiety, positive affect, defensiveness frequency, and stress were the strongest contributors to classification accuracy. High-conflict couples exhibited significantly higher levels of psychological distress, attachment insecurity, aggression, and dysfunctional communication behaviors, whereas low-conflict couples demonstrated greater emotional intimacy, relationship satisfaction, positive affect, and constructive conflict resolution patterns. The confusion matrix further indicated high sensitivity and specificity, confirming the robustness and generalizability of the classification model. The findings demonstrate that machine learning approaches can accurately distinguish high-conflict couples from low-conflict couples by integrating psychological, relational, and behavioral interaction variables. The results highlight the multidimensional nature of relationship conflict and suggest that relational functioning indicators and observed communication behaviors represent particularly powerful predictors of conflict status. Machine learning-based assessment frameworks may provide valuable tools for early identification, risk assessment, personalized intervention planning, and the development of data-driven approaches to couple therapy and relationship education.