A hybrid classification model for churn prediction based on customer clustering

Customer churn prediction is an active research topic for the data mining community and business managers in this rapidly growing society. The ability to detect churn customers precisely is something that every company would wish to achieve. From different experiments on customer churn, it can be seen that customers always could be divided into different types and the customers in the same segment generally have similar personas, behavioral preferences, and focus points. Therefore, a hybrid classification model named ClusGBDT for customer churn prediction is proposed. This model has three steps: a feature transformation stage, a customer clustering stage, and a prediction stage. At first, the multi-layer perceptron is used to training a prediction model and replace the original attributes with low-dimensional vectors. Then, customer segments are divided using K-means. Lastly, the unique prediction model based on GBDT is constructed for every customer segment. Several measures are used to evaluate the prediction performance. From the experiments, it is observed that our design could improve original classification algorithms include GBDT, random forest and logistic regression. Additionally, the proposed framework helps us to comprehend customer data.