Comparative analysis of explainable AI methods for telecom subscriber retention

DOI: 10.31673/2412-9070.2026.318114

Abstract

This paper presents a comparative analysis of Explainable Artificial Intelligence (XAI) methods applied to customer churn prediction in telecommunications companies. Subscriber churn is a contractual event preceded by behavioural signals occurring 14–90 days prior to contract termination or Mobile Number Portability (MNP) initiation, forming a distinctive pre-portability behavioural profile. A triaxial classification framework, TCC-XAI (Telecom Customer Churn XAI Framework), is developed along three dimensions: explanation granularity (D1), model dependency (D2), and practical applicability for retention (D3), supplemented by a fourth telecom-specific dimension of temporal awareness. Nine XAI methods are systematized within the framework and comparatively evaluated in terms of fidelity, stability, and actionability. TreeSHAP is shown to dominate at the descriptive-diagnostic applicability level, while DiCEML remains the only confirmed prescriptive method validated on telecommunications data. Three structural research gaps are identified: the absence of SHAP-versus-LIME stability benchmarking on large-scale telecom datasets; the absence of XAI analysis differentiated by subscriber service bundle type; and the absence of EBM benchmarking on churn prediction tasks. A three-layer operational TCC-XAI deployment pipeline for telecommunications operators is proposed. The scientific novelty of the study lies in the development of the first specialized TCC-XAI framework that simultaneously accounts for granularity, model dependency, practical applicability, and telecom-specific temporality, as well as in the identification of the multi-service gap as a critical unresolved research problem. The practical significance of the results consists in enabling the incremental deployment of XAI systems aligned with the operator's operational maturity, ensuring compliance with Article 22 of the GDPR, and improving the effectiveness of customer retention teams through the transition to prescriptive explanations.

Keywords: explainable artificial intelligence, customer churn, SHAP, LIME, counterfactual explanations, customer retention, subscriber retention.