Abstract: Facing the competition in the global market, enterprises are increasingly keener to explore how to hold the existing customers and improve their satisfaction by making use of the existing resources. The customer churn prediction has aroused more and more attention from enterprises. Given the unbalance and size of actual customer churn data, the paper puts forward an improved balanced-random forest algorithm and applies it to predict the customer churn of a commercial bank. The actual data set test result shows that the algorithm, on the strength of both sampling technique and cost-sensitive learning, has a higher accuracy in solving a large data set and unbalance data than the traditional prediction algorithms.

Key words: churn prediction, imbalanced data, random forests