Comparison of Two Main Approaches for Handling Imbalanced Data in Churn Prediction Problem

Customer churn is a major problem in several service industries such as banks and telecommunication companies for its profound impact on the company's revenue. However, the existing algorithms for churn prediction still have some limitations because the data is usually imbalanced. The commonly-used techniques for handling imbalanced data in churn prediction belong to two categories: resampling methods that balance the data before model training, and cost- sensitive learning methods that adjust the relative costs of the errors during model training. In this paper, we compare the performance of two data resampling methods: SMOTE and Deep Belief Network (DBN) against the two cost-sensitive learning methods: focal loss and weighted loss in churn prediction problem. The empirical results show that as for churn prediction problem, the overall predictive performance of focal loss and weighted loss methods is better than that of SMOTE and DBN.

Automated summary

Customer churn is a major problem for revenue in service industries like banks and telecommunication companies. Existing algorithms for churn prediction are limited by imbalanced data. This paper compares resampling methods with cost-sensitive learning methods and finds that the latter have better predictive performance in churn prediction.