This paper presents an explainable AI-driven framework for predicting net churn among prepaid mobile subscribers. The framework formulates churn prediction as a machine learning task mapped onto an industry reference architecture. Moreover, it introduces a quantile-based feature densification transform that reconciles learning features of unequal length and granularity into a homogeneous design matrix, applies a truncation heuristic that excludes low-confidence subscribers from binary classification to protect against label noise, and generates automated English-language insights describing the impact of each feature on churn propensity. We implement and validate the framework on real production data from a globally leading telco with tens of millions of prepaid subscribers spanning approximately 800 learning features and over 250 TB of commercial, experience, and network data. The truncation heuristic recovers a missed target classification performance at the second future inference period only at a cost of excluding roughly 3% of the subscribers. Our results suggest that subscriber commercial and experience indicators provide stronger churn signals than aggregated radio access network counters despite the greater representation of the latter in the data, and that churn risk rises as subscriber value declines, with high-value segments disproportionately exposed to gradual value leakage rather than outright attrition.

翻译：暂无翻译