1. Introduction
Retention campaigns are most cost-effective when targeted at subscribers genuinely at risk of churn, and interpretable models allow retention teams to design interventions matched to the specific factors driving an individual customer risk score.
2. Methodology
A dataset of 71,000 anonymised subscriber records with 34 behavioural and demographic features was used to train individual XGBoost, LightGBM and neural network classifiers, which were then combined in a logistic-regression stacking ensemble. SHAP (SHapley Additive exPlanations) values were computed on the ensemble to attribute prediction contributions to individual features.
3. Results
The stacked ensemble achieved an F1-score of 0.84 and AUC of 0.93 on a held-out test set, outperforming the best individual base learner (LightGBM, F1 0.81) by 3 points. SHAP analysis identified contract type, tenure and average monthly data usage as the three highest-magnitude features across the population, consistent with domain expectations from the retention team.
4. Conclusion
Ensemble stacking combined with SHAP-based interpretation delivers both predictive lift and actionable explanations for telecom churn management. Future work will incorporate call-centre transcript sentiment as an additional feature source.
References
[1] Lundberg S. M. and Lee S.-I., A unified approach to interpreting model predictions, NeurIPS, 2017. [2] Chen T. and Guestrin C., XGBoost: A scalable tree boosting system, KDD, 2016. [3] Ke G. et al., LightGBM: A highly efficient gradient boosting decision tree, NeurIPS, 2017.