Balancing Speed and Insight: Computational Efficiency of XAI Methods for Telecommunication Customer Churn Analysis

Abstract

As machine learning models become increasingly integral to business decision-making in telecommunications, the need for both transparent predictions and computationally efficient explanation methods has emerged as a critical challenge. This paper presents a rigorous cross-evaluation of four prominent explainable AI (XAI) methods-LIME, SHAP, ELI5 Permutation Importance, and native tree-based feature importance-applied to customer churn prediction models, with particular emphasis on their computational performance characteristics. We trained Random Forest and XGBoost classifiers on a telecommunications customer dataset (n=7,043) and conducted extensive benchmarking across varying dataset sizes and model complexities. Our results reveal significant disparities in execution time, with TreeSHAP for XGBoost demonstrating superior efficiency (0.0112s) compared to ELI5 with Random Forest (3.7114s)-a 331x performance difference. We quantify memory utilization patterns and establish scaling relationships between explanation time and both model complexity and dataset size. While computationally intensive methods like SHAP Kernel generally produced more comprehensive explanations, we found that optimized implementations like TreeSHAP offer an excellent balance between insight quality and computational efficiency. Notably, we observed substantial consistency in key churn predictors across methods, with contract type, tenure, and monthly charges consistently emerging as the most influential features. The paper concludes with a practical decision framework and technical implementation guidelines to help telecommunications providers select appropriate XAI methods based on their specific computational constraints, use cases, and interpretability requirements.