A Customer Churn Prediction Model Based on XGBoost and MLP

Abstract

Customer churn prediction is one of the central data mining tasks for researchers. In recent years, the customer churn prediction based on neural networks has achieved great success. However, traditional DNNs cannot achieve great predictive performance when facing numerical features. Hence, we proposed a hybrid prediction model based on XGBoost and multi-layer perceptron (MLP). Contrary to DNNs, the prediction model based on tree structure is adept in manipulating numerical features. This model has two stages. In the first stage, the XGBoost is applied to export the leaf number of customers based on numerical features. Then the MLP is used to deal with the one-hot vector transformed from the leaf number and the original discrete features in the second stage. The experimental results showed that our proposed model has a better predictive performance.