From prediction to intervention: Causal ML for explainable TV churn mitigation

Abstract

Customer retention is a significant challenge for Telecommunication Service Providers (Telecoms), which increasingly require transparent and explainable insights to prevent churn (subscription termination). Purely correlational models often behave as black boxes and may fail to reveal the underlying causes of churn or indicate how to reduce it. In contrast, causal methods aim to uncover root causes and, by quantifying the impact of targeted interventions, provide actionable guidance for churn mitigation. We study churn in Telenor Sweden’s TV services using 180K Internet Protocol TV (IPTV) and 24K coaxial TV customer records (Feb 2023–Jan 2024). We combine correlational prediction with causal discovery and interventional analysis to support proactive churn management when the true causal structure is unknown. For churn prediction under severe class imbalance, our EasyEnsemble-XGBoost model achieves weighted F1 scores of 89% (IPTV) and 91% (coaxial TV). For causal discovery, we propose a frequency-regularized ensemble that aggregates multiple state-of-the-art discovery algorithms into a consensus causal graph. In controlled simulations with known ground truth, our method outperforms baseline approaches in causal discovery (15.6% improvement in SHD), intervention identification (64.7% improvement in F1), and causal effect estimation (7.5% reduction in RMSE). On real TV churn data, the graph discovered by our method improves Bayesian Information Criterion (BIC) by nearly 4% over a Vote baseline. The estimated causal effects also pass multiple refutation tests. The resulting causal insights are action-oriented: increasing offers is estimated to reduce churn up to 34%, while SMS marketing can increase churn for certain customer segments up to 14%. © 2026 The Authors.