A Comparative Study of Metaheuristic Feature Selection and Optimized Deep Learning Models for Customer Churn Prediction in Banking

Abstract

Customer Churn Prediction plays a vital role in retaining customers in the highly competitive bank sector. The approaches used to preprocess data determine how effective deep learning and machine learning models are, even if they produce high-quality results, perform feature selection and optimize the model. This paper centers on the analysis of customer churn prediction through deep learning techniques using a dataset obtained from the Kaggle website. Data preprocessing with three steps like missing value imputation methods, data encoding, and normalization methods has been introduced for dataset. After preprocessing, data balancing is performed by Improved Synthetic Minority Oversampling Technique (ISMOTE) and Improved K-Nearest Neighbor (IKNN) is utilized to solve the data imbalance by giving minority groups the majority class on a regular basis. Feature Selection (FS) methods like Brownian Movement Dragonfly Optimization Algorithm (BMDOA), and Levy Flight Based Whale Optimization Algorithm (LFWOA) is applied to identify the most significant features for prediction. Convolution and Attention with Bi-directional Gated Recurrent Unit (CAT-BiGRU), Modified Weight and Activation Bidirectional Gated Recurrent Unit (MWABGRU), and Hyperparameter Tuned and Soft Activation Deep Multi-Layer Perceptron (HTSADMLP) have all been implemented to predict churn. The existing methods, Classification Time, and evaluation metrics like Area Under Curve (AUC), Kappa, Brier Score, and Expected Maximum Profit of the Customer (EMPC) are used to analyze the performance of the proposed method.