Source-linked AI summary
Explainable Artificial Intelligence for Customer Churn Prediction in Telecommunications: A Framework for CRM Integration
Sandeep Gaddamwar
TL;DR
Telecommunications churn is costly, and opaque predictive models leave retention staff without feature-level reasons for personalised interventions. This paper benchmarks four classifiers, evaluates SHAP and LIME explanations, and integrates their outputs into CRM workflows. Logistic Regression achieves an AUC-ROC of 0.8411, the leading models are effectively tied under cross-validation, and targeting the highest-risk quintile is projected to preserve $199K–$319K per campaign cycle while cutting overall churn by 3.3–5.3 percentage points.
Problem
Opaque churn models provide risk scores without principled feature-level reasons for selecting personalised retention interventions, while explanations are rarely connected to CRM workflows.
Method
The paper benchmarks four classifiers, applies global and instance-level SHAP and LIME explanations, and converts attribution vectors into CRM segmentation, retention actions, and feedback.
Results
Logistic Regression achieves an AUC-ROC of 0.8411, with tree-based alternatives within 0.011 AUC and the leading models effectively tied under cross-validation.
Takeaways & Limitations
The framework operationalises explanation outputs as CRM decision inputs, with top-quintile targeting projected to preserve $199K–$319K per campaign cycle and cut overall churn by 3.3–5.3 percentage points.
Takeaways & Limitations
Results rely on a single public benchmark lacking live-operator network-quality, competitive-pricing, and call-detail features, so performance and feature rankings may differ on proprietary data.
Abstract
from arXiv · showhide
Subscriber attrition is a costly, persistent challenge for telecommunications providers, with monthly churn of roughly 1.9% in mature markets eroding billions in revenue annually. Predictive models can flag at-risk customers accurately, yet they are routinely excluded from frontline CRM workflows because high-performing ensemble and non-linear architectures are opaque: a retention specialist cannot design a personalised intervention from a probability score alone, without knowing why a subscriber is at risk. This paper addresses that gap. We benchmark four classifiers--Logistic Regression, Random Forest, XGBoost, and LightGBM--on the IBM Telco Customer Churn benchmark (7,043 records; 19 features; 26.5% churn, balanced to 50% via SMOTE on the training partition only). Logistic Regression attains the strongest AUC-ROC (0.8411) and LightGBM the highest accuracy (78.42%); all four fall within a 0.011 AUC band (0.831--0.841), and 5-fold cross-validation confirms the leading models are effectively tied. Explanations are delivered at two granularities: a global SHAP ranking identifying tenure, total charges, and month-to-month contract as the dominant churn signals, and instance-level SHAP and LIME decompositions that expose the drivers behind each prediction. Building on these outputs, we introduce a four-layer CRM integration architecture that converts risk scores and attribution vectors into tiered segmentation, maps top features to structured retention-action templates, and routes campaign outcomes into a retraining feedback loop. Targeting the highest-risk quintile is projected to cut overall churn by 3.3--5.3 percentage points, preserving an estimated $199K--$319K per campaign cycle.
I. INTRODUCTION
Telecommunications churn is financially costly, while opaque predictive models limit personalised retention decisions. This work combines classifier benchmarking, explainability methods, and CRM integration to turn predictions into operational retention workflows.
- 1.9% monthly churn at U.S. wireless carriers creates substantial annualised revenue erosion at industry scale.
- Opaque ensemble models provide churn scores without showing which features justify contract offers, bill credits, or outbound calls.
- GDPR Article 22 and EU AI Act transparency obligations increase pressure for meaningful explanations of automated decisions.
- Prior XAI work rarely specifies how SHAP or LIME outputs should become CRM actions or feed campaign outcomes back into models.
- The study benchmarks four classifiers, combines global and local explanations, and proposes a CRM architecture using attribution-based action templates.
- $199K–$319K per campaign cycle is projected to be preserved by targeting the highest-risk quintile, reducing overall churn by 3.3–5.3 percentage points.
II. RELATED WORK
Related work established classifiers for churn prediction, post-hoc explanation methods, and predict-segment-act retention management. However, the literature has not fully connected explanation outputs to explicit CRM actions and business-value-oriented intervention.
- Logistic regression provided an interpretable churn baseline, while Random Forest and gradient boosting enabled modelling of more complex interaction effects.
- SHAP uses additive Shapley-based attributions, whereas LIME fits a locally weighted linear surrogate prioritising local fidelity.
- Retention research frames operations as predict-segment-act: score churn propensity, assign risk tiers, and route customers to predefined programmes.
- Intervention effectiveness depends on identifying persuadable customers, and maximising AUC does not necessarily maximise business value.
D. The Gap This Work Addresses
The paper addresses the gap between analytical explanations and operational CRM use. It combines benchmark data preparation with an architecture that converts model attributions into structured retention decisions.
- Existing churn studies usually present SHAP or LIME charts without specifying how attribution vectors become personalised retention offers or model feedback.
- The proposed architecture treats explanation outputs as first-class CRM inputs rather than data-science-only analytical deliverables.
- 7,043 benchmark records and 19 predictive features define the IBM Telco Customer Churn evaluation setting.
- SMOTE is applied only to the training partition, producing 8,278 balanced training records while preserving an uncontaminated 1,409-record test set.
C. Models
The study compares four classifiers and applies SHAP and LIME at global and instance levels. Its explanation pipeline measures feature influence, decomposes individual predictions, and supports local surrogate analysis.
- Logistic Regression, Random Forest, XGBoost, and LightGBM are trained and compared, with logistic regression serving as the interpretable linear baseline.
- Global SHAP importance ranks features by mean absolute attribution across test instances, identifying consistently influential predictors.
- Permutation importance provides a model-agnostic cross-check that agrees on the two dominant predictors.
- Instance-level SHAP attributions sum to the gap between predicted log-odds and the dataset base value, with positive values pushing toward churn.
- LIME discretises continuous variables and treats binary and one-hot columns categorically to produce instance-specific local surrogates.
E. CRM Integration Design
The framework combines subscriber risk signals, model performance evidence, and CRM-oriented inputs to support churn segmentation and intervention design.
- Risk signals: 42.7% of month-to-month subscribers churn, compared with 11.3% on one-year and 2.8% on two-year contracts.
- Risk signals: 10 months is the median tenure among churners, versus 38 months among retained customers; mean monthly charges are $74.44 and $61.27, respectively.
- Model evaluation: 0.8411 AUC-ROC is achieved by Logistic Regression, while LightGBM records 78.42% accuracy across the compared classifiers.
- Risk signals: The EDA overview compares class balance and churn rates by contract, internet-service type, and payment method.
C. Global Feature Importance
Global SHAP analysis identifies tenure as the strongest churn signal, followed by total charges, contract type, fiber service, and paperless billing.
- Tenure has more than 2.5× the mean absolute SHAP attribution of any other feature, making it the strongest churn signal.
- TotalCharges ranks second and closely tracks tenure because the two features are mechanically correlated.
- Month-to-month contracts elevate churn attribution, whereas two-year contracts are strongly protective.
- Fiber-optic service and PaperlessBilling are subsequent churn-elevating predictors, consistent with the exploratory churn patterns.
D. Individual Explanations: Two Case Studies
Instance-level SHAP and LIME explanations connect predicted churn probabilities to subscriber-specific risk and protective factors, supporting operational interpretation.
- Case studies: 0.941 is the predicted churn probability for a new subscriber on a month-to-month contract.Short tenure is the dominant risk-elevating feature, while low total charges partially offset the prediction.
- Case studies: 0.003 is the predicted churn probability for a retained subscriber with long tenure and a two-year contract.These protective features route the subscriber toward low-priority treatment.
- Dependence plots: Fig. 6 plots the three highest-importance features against their values, colored by predicted churn probability.Tenure and total charges show monotonic effects, while contract type separates month-to-month from committed subscribers.
- Model diagnostics: Fig. 2 pairs ROC curves for all classifiers with confusion matrices evaluated at the 0.5 threshold.
- CRM integration: The architecture sends risk-tier assignments, SHAP attribution vectors, and LIME coefficient vectors into the CRM action pipeline.
B. Segmentation
The segmentation design assigns subscribers to risk tiers and uses attribution-driven action templates, with campaign outcomes feeding action optimization and retraining.
- Risk tiers: High risk is defined as p > 0.70, Medium as 0.40 ≤ p ≤ 0.70, and Low as p < 0.40.
- Risk tiers: 26.0% of the test set falls into High risk, 23.2% into Medium risk, and 50.8% into Low risk.
- Calibration caveat: SMOTE-balanced training inflates output probabilities relative to the 26.5% deployment prevalence, requiring recalibration before literal threshold interpretation.
- Action mapping: Top-attribution features map to structured interventions such as contract upgrades, service reviews, payment incentives, and tech-support trials.
- Feedback loop: Campaign outcomes are logged with scores and attribution vectors so uplift statistics can optimize action assignments and support retraining.
E. Business Impact
Targeting the highest-risk quintile is projected to reduce overall churn while preserving substantial revenue per campaign cycle. The paper also argues that near-parity among classifiers makes the simpler logistic model practically relevant for this benchmark.
- $199K–$319K per campaign cycle is projected to be preserved by targeting the top 20% of at-risk subscribers.The estimate assumes campaign uptake rates of 25%–40% among genuine churners reached and a 12-month customer lifetime value of $850.
- 3.3–5.3 percentage points is the projected reduction in overall churn from targeting the highest-risk quintile.The figures are illustrative and depend on assumed uptake and customer-lifetime-value parameters.
- Near-parity among the four classifiers suggests that a regularised logistic model captures the largely additive churn signal efficiently in this account-level benchmark.The paper cautions that richer sequential, network, or call-detail data may favour more complex models.
B. Performance in Context
The reported AUC-ROC results fall within the range found in prior IBM Telco benchmark studies, but accuracy comparisons depend on class balance and evaluation setup. Practically, the framework emphasizes targeted contract management, monitoring for distribution shift, and caution when interpreting correlated-feature attributions.
- B. Performance in Context: 0.831–0.841 is the paper’s AUC-ROC band, within the range reported for the IBM Telco benchmark literature.Accuracy figures above roughly 85% evaluated on the natural 26.5% churn imbalance are not directly comparable to this SMOTE-balanced, recall-favouring operating point.
- C. Implications for Practitioners: Contract-term management is identified as the highest-leverage intervention, with SHAP-to-action mappings supporting operational use by CRM teams.The proposed action is converting month-to-month subscribers to one-year contracts before churn probability rises.
- C. Implications for Practitioners: PSI monitoring is intended to flag model obsolescence when the subscriber base shifts because of acquisitions, pricing changes, or new products.This provides a circuit-breaker against silent performance degradation.
- C. Implications for Practitioners: SHAP waterfall plots show how individual features move predictions in log-odds space, with red increasing churn risk and blue increasing retention.The examples contrast a high-risk subscriber at p = 0.941 with a retained subscriber at p = 0.003.
- D. Limitations: The benchmark uses a single public account-level dataset, while correlated predictors, assumed business-impact parameters, and cross-sectional data constrain interpretation and generalisation.The paper states that attributions are associational rather than causal and that proprietary or sequential data may produce different results.
E. Ethical Considerations
The paper identifies ethical risks in using demographic proxies for churn prediction, especially potential differential treatment through omitted retention offers. It therefore calls for fairness audits and privacy-conscious feature governance alongside CRM integration.
- Demographic proxies such as senior-citizen, partner, and dependent indicators may encode differential treatment across population groups.
- Fairness audits for statistical parity and equalised odds across age, gender, and family-structure groups are presented as operational necessities before deployment.
- Privacy governance should apply data minimisation to subscriber information assembled from billing, network, and support systems.
- Features whose marginal predictive value does not justify their privacy cost should be excluded.