Federated Machine Learning based Bank-Customer Churn Prediction

Abstract:

Customer Churn Prediction is a strategy used by banks to search and identify customers who show a high tendency to leave the company. The banking industry faces challenges to hold its customers, and it is an important part of customer-oriented retention that aims to reduce churners as there are around 1.5 million churn customers in a year, a number that is increasing every year. Customer retention has a prodigious impact on the bottom line of a bank’s profits, and this kind of impact has exceeded that caused by scale, market share, unit cost, and other relevant factors of competitive advantage. Due to multiple data breaches from banks in the past few years, privacy has become a critical requirement. With the aim of guaranteeing user privacy and data confidentiality, we propose the idea of Federated Machine Learning or FedML to train a Multi-layer Perceptron (MLP) model at each branch of a bank where a data source resides. These models collaboratively train a global model located at the central server. So, the sensitive data is not shipped to the central server and the training process is carefully engineered so that no site can speculate the private data. At the same time, the centralized model is built as if the data sources were combined. This results in a lesser computational load on the central server. We compare our method with the standard approach of server training to prove the robustness of our method.

Practical Data Mining Techniques and Applications