Comparison of XGBoost LightGBM and CatBoost for Online Store Customer Churn Prediction

Abstract

Customer churn is a critical issue in digital business because it can reduce revenue, increase customer acquisition costs, and weaken a company’s competitive position. In the context of online stores, customers have many alternative shopping platforms, requiring companies to understand customer behavior patterns to minimize churn risk. This study aims to develop an online store customer churn prediction model using tree-based gradient boosted models by comparing XGBoost, LightGBM, and CatBoost algorithms. This study applies an experimental method using a secondary dataset obtained from Kaggle. The dataset consists of 25,000 observations and 13 variables related to online store customer transactions and characteristics. The research stages include problem identification, data collection, preprocessing, missing value handling, removal of irrelevant variables, data visualization, outlier detection, feature engineering, model selection and training, model evaluation, and result analysis. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC curve. The results show that XGBoost achieved the best performance with an accuracy value of 0.80032 and an ROC-AUC value of 0.66. Feature importance analysis indicates that the Product_Category_Clothing variable was the most influential feature in the model performance. These findings show that tree-based gradient boosted models, particularly XGBoost, can be used to help online store companies identify customers who are likely to churn and design more targeted customer retention strategies.