Explainable AI Turns Telecom Churn Predictions Into Actionable Retention Strategies

Customer churn remains one of the most expensive problems in the telecommunications industry. Every subscriber who quietly cancels a contract takes recurring revenue with them, and replacing that customer typically costs far more than retaining the original one. Machine learning models have long been able to flag which customers are likely to leave, but a persistent criticism has haunted these systems: they are black boxes. A model might predict that a specific subscriber is about to churn with impressive accuracy, yet offer no explanation of why, leaving retention teams unable to act on the warning. A new study published in Discover Sustainability by researchers at Parul University in Vadodara, India, tackles this gap head-on with a dual layer explainable artificial intelligence framework that not only predicts churn but translates each prediction into a concrete, business-ready action.

The research team, led by Snehlata Barde together with Sneha Soni, Dinesh Swami and Kaustubh Sutar, built their framework around the widely used Telco Customer Churn dataset, a benchmark collection of subscriber records containing contract details, service subscriptions, billing information and demographic attributes. At the prediction layer, the authors deployed XGBoost, a gradient boosted decision tree algorithm renowned for its performance on structured tabular data. XGBoost works by building an ensemble of decision trees sequentially, with each new tree trained to correct the residual errors of its predecessors. This iterative error-correction mechanism, combined with regularization terms that penalize model complexity, allows the algorithm to capture subtle nonlinear interactions between features, such as the way month-to-month contracts combined with fiber optic service and electronic billing can jointly elevate churn risk in ways no single feature reveals on its own.

Accuracy alone, however, was never the end goal. The second layer of the framework is where explainable artificial intelligence enters. The researchers applied two complementary explanation techniques, SHAP and LIME, to illuminate the reasoning behind every individual churn prediction. SHAP, which stands for SHapley Additive exPlanations, draws on concepts from cooperative game theory to distribute credit for a prediction across all input features. By treating each feature as a player in a game and the prediction as the payout, SHAP computes Shapley values that quantify exactly how much each attribute pushed a given prediction above or below the average. This approach carries desirable theoretical properties, including consistency and local accuracy, which make the attributions mathematically trustworthy rather than merely plausible.

LIME, or Local Interpretable Model-agnostic Explanations, takes a different route to the same destination. Instead of attributing credit through game-theoretic enumeration, LIME probes the neighborhood of a single prediction by perturbing the input slightly and observing how the model responds. It then fits a simple, interpretable surrogate model, typically a sparse linear model, to those perturbed samples, so that the surrogate faithfully reproduces the behavior of the complex XGBoost model in the immediate vicinity of the customer in question. The coefficients of that surrogate reveal which features matter most for this particular subscriber. Where SHAP offers theoretically grounded global-to-local attributions, LIME offers fast, intuitive local approximations, and running both provides a form of cross-validation for the explanations themselves.

That cross-validation is not left to intuition. The study quantifies the agreement between SHAP and LIME using two statistical measures: Spearman’s correlation coefficient, which assesses whether the two methods rank feature importance in the same order, and the Jaccard index, which measures the overlap between the sets of top features each method identifies. When two fundamentally different explanation techniques converge on the same story for a given customer, confidence in that explanation rises substantially. When they diverge, analysts know to scrutinize the prediction more carefully before committing retention resources. This dual-framework validation addresses one of the most common objections to explainable AI, namely that a single explanation method might be flattering the model rather than genuinely exposing its logic.

The most novel contribution of the paper is what the authors call the Strategic Actionability Score, or SAS. Explanations, after all, are only useful if someone can do something with them. The SAS takes the feature attributions produced by SHAP and LIME and evaluates them through a business lens, determining the strategic value of acting on each prediction. A churn prediction driven by a modifiable factor, such as a customer’s contract type or an unnecessary add-on service, scores higher on actionability than one driven by an immutable attribute. This scoring mechanism effectively converts the output of a statistical model into a prioritized worklist for retention teams, telling them not just who is at risk but which levers are worth pulling for each individual subscriber.

The implications reach beyond marketing efficiency. The authors frame their framework as a contribution to sustainable customer relationship management, arguing that transparent, explanation-driven retention strategies reduce unnecessary operational expenditure and support efficient resource utilization. Instead of blanket discounts offered indiscriminately to an entire at-risk segment, telecom providers can target interventions precisely where the explanations indicate they will matter, cutting waste and improving long-term operational sustainability. In an industry where acquisition costs routinely exceed retention costs by a wide margin, the economic case for precision retention is compelling, and the environmental case for reducing redundant outreach and infrastructure strain adds a dimension rarely discussed in churn analytics.

The work also speaks to a broader movement in artificial intelligence toward responsible and accountable deployment. Regulatory pressures, including data protection regimes that grant individuals rights to explanation in automated decisions, are pushing industries away from opaque models regardless of their accuracy. A framework that pairs a high-performing predictor with mathematically grounded explanations, validates the agreement between two independent explanation methods, and distills the results into an actionability score offers a template that extends well beyond telecommunications. Banking, insurance, subscription media and any sector facing customer attrition could adapt the same dual layer architecture, swapping the dataset and retraining the XGBoost model while retaining the explanation and actionability machinery.

Limitations and future directions remain. The study relies on a single benchmark dataset, and real-world telecom data is messier, higher-dimensional and constantly shifting as services and pricing structures evolve. Explanation agreement metrics like Spearman’s correlation and the Jaccard index provide useful signals but do not guarantee causal insight; a feature that correlates with churn risk is not necessarily a lever that causes churn to rise or fall when manipulated. The authors acknowledge that the framework’s value lies in guiding responsible, transparent strategies rather than replacing human judgment, and the field will need longitudinal studies to confirm that explanation-guided interventions actually outperform intuition-led ones in live deployments.

Nevertheless, the study marks a meaningful step in the maturation of applied machine learning. The era of celebrating raw accuracy is giving way to an era in which models must justify themselves, and in which the justification must be checked, quantified and converted into action. By wiring SHAP and LIME into a validation loop and capping the pipeline with a business-oriented actionability score, the Parul University team has demonstrated that the distance between a prediction and a prudent decision can be shortened to a single, auditable workflow. For telecom operators watching subscribers slip away, and for the growing community of researchers insisting that artificial intelligence explain itself, that distance is precisely where the future of customer analytics is being decided.

Subject of Research: Explainable AI for telecom customer churn prediction and actionable retention analytics

Article Title: A dual layer explainable artificial intelligence approach for prediction and action in telecom customer churn analytics

Article References: Barde, S., Soni, S., Swami, D., & Sutar, K. (2026). A dual layer explainable artificial intelligence approach for prediction and action in telecom customer churn analytics. Discover Sustainability. https://doi.org/10.1007/s43621-026-04867-2

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04867-2

Keywords: dual, layer, explainable, artificial, intelligence, approach, prediction, action, telecom, customer, churn, analytics