A PCA-AdaBoost model for E-commerce customer churn prediction

Customer churn is detrimental to corporate revenue. Hence, accurate customer churn prediction is vital for enterprises to improve customer retention and corporate revenue. However, in e-commerce, there are challenges when predicting customer churn using traditional models. First, the conventional recency, frequency, and monetary (RFM) analysis based on single-source data can hardly accurately predict the e-commerce customer churn since such customers are often non-contractual. Second, in general, the data about e-commerce customers is high-dimensional and unbalanced, making traditional models even less effective. In this paper, we propose an improved prediction model by integrating data pre-processing and ensemble learning to solve these two problems. Specifically, two new features are first integrated into the RFM analysis to better capture customer behaviors. Second, the principal component analysis is adopted to reduce data dimensions. Third, adaptive boosting (AdaBoost) is employed to cascade multiple decision trees to minimize the impacts from the unbalanced data. For clarity, this model is called the PCA-AdaBoost model. We use an e-commerce dataset published on the Kaggle platform to demonstrate its effectiveness by conducting numerical experiments. We compare the performance of the PCA-AdaBoost model developed in this paper with several models proposed in the literature. Our results confirm that the PCA-AdaBoost model can achieve more accurate customer churn prediction and better overall stability.