Abstract:
"Customer churn also known as customer attrition is a phenomenon where customers or subscribers of a service, or product cease their relationship with the organization. Customer churn can occur for various reasons, including dissatisfaction with the product or service, better offers from competitors, or changes in personal circumstances. Measuring and understanding churn is crucial for businesses and organizations as it directly impacts revenue and profitability. High churn rates signal underlying issues with customer satisfaction, product or service quality, or market competition. The banking domain is among the many sectors that face customer churn. Annually, about 1.5 million bank customers churn, posing significant losses for the institutions. Acquiring new customers can be costly and establishing loyalty takes time. Hence early churn prediction has become an area focus for banks in taking measures to maximize the retention periods of their customers.
This research study presents a novel deep-learning ensemble model to predict customer churn early on in the banking domain. The proposed model consists of four stages. Initially, three key data preprocessing techniques are employed, followed by the implementation of a feature reduction system using XGBoost. In the third level, four distinct classifiers including CNN, SVM, LR, and DT are trained as base classifiers. Finally, a stacking ensemble method combines the predictions from these base classifiers, utilizing CatBoost as a meta-learner to exploit the strengths of each classifier.
The proposed system obtained positive results for the proposed algorithm. The ensemble model outperformed the individual classifiers in terms of accuracy and performance. The proposed model obtained 89.25% of accuracy, 85.60% of precision, 54.45% of recall rate, an F1-score of 0.6656, an AUC value of 0.7610, and a Kappa value of 0.6053. Further, the prototype model was benchmarked against the existing ensemble models and achieved higher results in terms of accuracy surpassing all the similar models’ results."
