Logistic Regression Reaches 0.8411 AUC-ROC on IBM Telco Churn with Instance-Level SHAP

The study benchmarked logistic regression, random forest, XGBoost, and LightGBM on the IBM Telco Customer Churn set after SMOTE balancing of the training partition. Logistic regression led AUC-ROC at 0.8411 while LightGBM reached 78.42 percent accuracy; all models stayed within a 0.011 AUC band. Global SHAP identified tenure, total charges, and month-to-month contracts as dominant signals. Instance-level SHAP and LIME decompositions supplied per-subscriber feature attributions for retention specialists.

The four-layer integration converts risk scores and attribution vectors into tiered segmentation, maps top features to action templates, and feeds campaign outcomes back into retraining. Targeting the highest-risk quintile yields the stated revenue preservation range of $199K-$319K per cycle. This architecture directly addresses the documented exclusion of opaque ensembles from frontline CRM systems.

Prior churn studies focused on accuracy alone or used post-hoc explanations without operational mapping. The present work closes that gap by specifying how attribution vectors route into structured interventions and closed-loop retraining. Regulatory pressure on automated decisions in telecommunications increases the requirement for auditable feature contributions over raw probability outputs.

Legacy CRM platforms already ingest structured customer fields; the proposed templates align directly with existing contract and billing attributes. Real-time scoring pipelines can attach SHAP vectors without additional latency once feature stores are extended. Deployment records from similar XAI layers in billing fraud systems show sustained lift when feedback loops update models quarterly.