Counterfactual Ensembles for Interpretable Churn Prediction: From Real-World to Privacy-Preserving Synthetic Data

Counterfactual explanations identify minimal input changes needed to alter a machine learning model’s prediction, offering actionable insights in tasks like churn analysis. However, existing methods often produce counterfactuals that vary in quality, coherence, and plausibility, limiting their practical value. We propose an ensemble evaluation framework that integrates multiple generation techniques and ranks their outputs using a tunable scoring function balancing multiple relevant metrics. Our approach addresses two key deployment scenarios: (i) in-house churn analysis, where decision-makers can interactively adjust scoring weights for tailored, user-driven explanations; and (ii) outsourced churn prediction, where counterfactuals must be generated on synthetic data to preserve privacy while remaining representative of real cases. Experiments on benchmark churn datasets demonstrate that our ensemble approach improves the consistency, interpretability, and utility of counterfactuals across both real and synthetic settings, supporting more reliable and privacy-aware decision-making.