Optimized customer churn prediction using tabular generative adversarial network (GAN)-based hybrid sampling method...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced and overlapped data in customer churn prediction negatively impact classification accuracy.
• Existing sampling and hybrid sampling methods show limitations with classical machine learning algorithms.
• Classical algorithms struggle with imbalanced datasets common in churn prediction.

Purpose of the Study:

• To optimize classical machine learning performance for customer churn prediction.
• To introduce CostLearnGAN, a cost-sensitive learning framework integrating generative adversarial networks (GANs) and hybrid sampling.
• To enhance the effectiveness of classical machine learning algorithms in identifying customer churn.

Main Methods:

• Developed CostLearnGAN, a tabular generative adversarial network (GAN)-hybrid sampling method.
• Applied a cost-sensitive learning perspective to improve classical machine learning algorithms.
• Conducted experiments using six comparative sampling methods, six datasets, and three machine learning algorithms.

Main Results:

• CostLearnGAN achieved superior performance across all evaluation metrics, with an average mean rank score of 1.44.
• Demonstrated CostLearnGAN's robustness, outperforming other sampling methods with an average robustness value of 5.68.
• Validated the efficiency of classical machine learning algorithms with shorter execution times for large-scale churn prediction.

Conclusions:

• CostLearnGAN significantly improves the performance and robustness of classical machine learning models in customer churn prediction.
• The cost-sensitive learning approach combined with GAN-hybrid sampling offers an effective solution for imbalanced data challenges.
• Classical machine learning algorithms, enhanced by CostLearnGAN, are suitable for efficient churn prediction in large customer bases.