Artificial Neural Network Using Weight Initialization in Customer Churn Prediction: Banking Industry

Predicting customer churn is crucial but ANN have often underperformed compared to other models in the competitive banking industry. This study focuses on enhancing an Artificial Neural Network (ANN) model by exploring the effects of various weight initialization techniques, batch normalization, and dropout layers. Our findings indicate that the LeCun Normal initializer, when combined with batch normalization and dropout, achieved the highest validation accuracy of 84.17%, representing a significant improvement over the default Glorot Uniform as the default Keras initializer. We also found Glorot Normal and He Uniform also showed strong and consistent result with accuracies up to 83.94% and 83.87%. On the other hand, the Zero initializer consistently underperformed. These findings highlight the importance of choosing the right initializers and regularization techniques to improve ANN performance in churn prediction.