Ensemble based Machine Learning Approach for Predicting Customer Churn in E-Commerce

Due to the intense competition among E-Commerce industries, customers have a large selection of products to choose from and as a result, these industries experience customer churn. Since retaining existing customers than acquiring new ones is cost effective, customer churn prediction is key for the E-commerce industry. The goal of this paper is to forecast customer churn for a ride sharing application by training appropriate supervised ensemble-based machine learning models including Gradient Boost, Cat Boost and XG Boost on a known dataset containing 50,000 known samples of many such customer attributes as well as their loyalty to the application. A variety of manual preprocessing approaches, such as removing missing data, duplicate data, outliers, etc., were employed to reduce the high-class imbalance in the raw dataset while also preventing useful information loss in the already small dataset, enabling the models to predict customer churn with neutrality. Standardization techniques was also utilized to normalize the data so as to restrict the wide distribution of each feature and improve model performance. The AUC-ROC performance metric was used, which met our objectives while also working well with imbalanced datasets. Gradient Boost and Cat boost algorithms performed the best in identifying potential churning customers, with AUC-ROC scores of 0.85 and computation times of 58 seconds and 80 seconds respectively, while XG boost was not far behind with an AUC-ROC score of 0.84 and the fastest computation time of 45 seconds among all models assessed.