Customer Churn Prediction: Performance and XAI (SHAP-LIME) Analysis

Abstract

With the advancement of digital transformation, the increasing volume of data and intensifying competition have made customer churn a significant risk in the telecommunications sector. Therefore, data-driven approaches for predicting customer churn have become increasingly important. However, evaluating these approaches solely on model performance is insufficient; it is also essential that the models' decision-making processes be understandable. This study aims to evaluate churn prediction models in terms of both performance and interpretability. Within this scope, a dataset obtained from Kaggle was analyzed for missing values and outliers, and data transformations were applied, followed by the SMOTE method to address class imbalance. Models were developed using Logistic Regression, Random Forests, Decision Trees, Naive Bayes, and K-Nearest Neighbors algorithms, and their performance was evaluated using performance metrics. To understand the model's behavior and the effects of variables, the SHAP method was applied at the global level, while SHAP and LIME were used at the local level to explain individual predictions; the results from SHAP and LIME were then compared. Therefore, explainable artificial intelligence methods can enhance the transparency and reliability of model outputs, thereby making significant contributions to decision support processes.