What question did this study set out to answer?

The research aims to predict customer churn in telecommunications using an explainable AI framework.

February 13, 2026Open Access

Explainable AI-driven customer churn prediction: a multi-model ensemble approach with SHAP-based feature analysis

Key Points

The research aims to predict customer churn in telecommunications using an explainable AI framework.
Utilized the Telco Customer Churn dataset with 7,043 records.
Implemented comprehensive feature engineering and SMOTE oversampling.
Trained seven machine learning models, including XGBoost and Random Forest.
Conducted model interpretation through SHAP analysis and customer segmentation.
Performed threshold optimization to improve precision and recall.
Gradient boosting models achieved high performance, with the best F1-score of 0.84.
XGBoost had the highest discriminative ability indicated by an AUC-ROC of 0.932.
A soft-voting ensemble of top models matched the best performance with an AUC-ROC of 0.918.
Key features for churn predictions included contract type, tenure, and technical support.
Threshold optimization improved precision to 0.90 and recall to 0.91 while reducing false negatives by 15%.

Abstract

Customer churn prediction is critical for telecommunications companies to maintain profitability and inform retention strategies. This study builds upon existing work by implementing a comprehensive machine learning framework using the Telco Customer Churn dataset ( n = 7,043). Our methodology integrated comprehensive feature engineering, SMOTE oversampling, and training of seven machine learning models including XGBoost, Random Forest, and a Multi-layer Perceptron. Model interpretation was conducted via SHAP analysis and customer segmentation. Key results demonstrated that gradient boosting algorithms (XGBoost, LightGBM, Gradient Boosting) achieved the highest balanced performance with accuracy, precision, recall, and F1-scores of 0.84, with XGBoost attaining the best discriminative ability (AUC-ROC: 0.932). A soft-voting ensemble of the top models matched this performance (F1-score: 0.84, AUC-ROC: 0.918). SHAP analysis revealed that contract type, tenure, and technical support were the features contributing most to the model's churn predictions. Threshold optimization at 0.528 balanced precision (0.90) and recall (0.91) while reducing false negatives by 15%. The findings provide actionable insights for prioritizing high-risk customers and designing targeted retention strategies in the telecom sector.

Perguntar à IA

Bookmark

View Full Paper