Actionable and Interpretable ML-Based Early Warning Systems for Divorce Incorporating Causal Inference and Counterfactuals

We present an interpretable machine learning framework for divorce prediction that integrates causal inference and counterfactual reasoning to generate actionable insights. Using a dataset of 170 couples (85 divorced, 85 married) assessed via the Divorce Predictors Scale (54 behavioral features), we identify 16 causally significant predictors using Double Machine Learning with Causal Forests. Notable drivers include humiliating language during arguments (ATE: +24.4%) and shared entertainment preferences (ATE: –20.8%). We train four gradient-boosting models—XGBoost, LightGBM, CatBoost, and HistGradientBoosting—and achieve high performance, with XGBoost yielding 97.9% accuracy and CatBoost achieving a ROC-AUC of 0.99. Our models outperform prior approaches, including a BERT-based Random Forest (accuracy: 81.0%) and also outperforms the state-of-theart transformer model (FT-Transformer accuracy: 97.0%), while providing greater interpretability. Our framework uniquely combines SHAP values for local and global explanations (e.g., humiliation contributing 0.11 units toward an individual’s divorce prediction), DiCE for generating diverse and plausible counterfactuals (e.g., reducing humiliation flipped the prediction to marital stability), and Bayesian Neural Networks to estimate uncertainty (±9.3% standard deviation). The entire system is accessible via an opensource Google Colab notebook, allowing users to simulate personalized interventions. This research contributes to the fields of responsible AI, computational and informational social science by demonstrating that high-accuracy prediction can be meaningfully combined with interpretability and actionability in sensitive domains. All reproducibility resources are provided via a publicly available GitHub repository