Customer Churn Prediction System using Deep Learning Techniques

Abstract

Modern business operations necessitate churn prediction systems, particularly in industries with subscription-based services and recurring revenue streams. This study proposes a deep learning-based customer churn prediction system for e-commerce, aiming to identify and retain at-risk customers. Addressing the limitations of previous studies—such as sparse representation of industry-specific challenges and limited exploration of deep learning models and feature engineering—the study utilizes the cell2cell dataset to build predictive models. Techniques like principal component analysis, correlation matrices, and random forest classifiers, alongside Python tools, were employed for data preparation and feature selection. Advanced deep learning methods, including recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, were used to improve prediction accuracy and capture temporal relationships in customer behavior. Performance was evaluated using metrics like accuracy, precision, recall, F1 score, and AUC score. Results indicated that RNN achieved accuracy and AUC scores of 0.7108 and 0.6107, respectively, while LSTM achieved scores of 0.7106 and 0.6026. The study recommends future research to employ deep learning and advanced feature engineering approaches to enhance churn prediction systems.