Addressing Silent Churn in a Non-Contractual Setting : An intelligent decision support system integrating machine learning, explainable artificial intelligence and large language models

Sammanfattning

Shared mobility is capital-intensive and every retained customer protects margins. However, customer churn in non-contractual services is difficult to manage because customers leave without notice. Additionally, existing research tends to address churn prediction in isolation, leaving a gap between identifying at-risk customers and acting on that knowledge. In response to this gap, this study develops an Intelligent Decision Support System to address silent churn in a station-based, non-contractual car-sharing context. The system operates in three sequential layers: an XGBoost model identifies customers at risk, SHAP and K-Prototypes clustering surface the behavioral drivers behind those predictions, and a large language model translates these drivers into targeted retention strategies. The system is evaluated on data from a car-sharing operator in Sweden, covering over 23 000 customers and 350 000 trips. The prediction phase achieves a PR-AUC of 0.87 on the held-out test set, with a false alarm rate of approximately 6.5%. The explanation phase divides at-risk customers into seven distinct behavioral personas, and in the retention phase each persona receives targeted strategies delivered in a business-ready report, assessed through expert evaluation at the case company. The findings suggest that combining segment-adjusted churn definitions with an integrated analytical pipeline produces a system that is both predictively strong and practically useful. The study contributes to the literature by addressing the non-contractual churn definition problem and the methodological gap between prediction, explanation and retention within a single integrated framework.