Authors: Vasileios Gkonis, Ioannis Tsakalos

Title: Deep Dive into Churn Prediction via Deep Learning in Banking Sector: The challenge of Hyperparameters selection and imbalance learning.

Abstract

Customer churn has been one of the most major issues in Banking sector over the years. Customers may strive to switch financial institutions to gain more as a result of better financial services and products. Hence, the early identification of customer's leaving is crucial for the sustainability of Banking institutions. However, Churn modeling is hampered by imbalanced datasets between classification classes, where the churn class is typically significantly smaller. In this Study we seek to examine the performance of Deep Learning in bank churn prediction, while incorporating various sampling techniques to address the challenges posed by imbalanced datasets. For this purpose, we examine the performance of Synthetic Minority Oversampling Technique (SMOTE) and two mixed sampling techniques SMOTE-Tomek Links and SMOTE-Edited Nearest Neighbors (SMOTE-ENN). Over the last years Deep Neural networks(DNN) have demonstrated significant predictive power in a variety of tasks and industries. Nevertheless, as a result of their architecture complexity, the proper selection of training hyperparameters is crucial. In this study we evaluate the performance of several Activation Functions and Optimizers, which are critical components for the DNNs performance. ReLu, Swish, Mish and APTx activations functions have been tested and Adam and R-Adam as optimizers. We explore the effectiveness of different activation functions and optimizers in combination, aiming to identify configurations that yield promising results for churn prediction tasks. Our results shown that mixed resampling strategies may yield to better results than Oversampling Techniques. In addition, the performance of APTx activation function which has recently proposed has shown promising results. Regarding the Optimization process R-Adam seems to achieve better performance when compared to Adam. The insights of this study aim to aid researchers and practitioners in deploying effective Deep Learning Models for Churn Prediction in Banking sector.