Research on customer churn prediction and model interpretability analysis.

Area of Science:

• Computational Intelligence
• Financial Technology
• Machine Learning Applications

Background:

• Intensified competition and diversified customer choices in the banking industry lead to increased customer churn.
• Predicting customer behavior and retaining existing customers are critical challenges for commercial banks.
• Traditional customer loyalty is declining due to advancements in information and internet technologies.

Purpose of the Study:

• To construct an accurate customer churn prediction model for commercial banks using bank business data.
• To identify key features influencing customer churn through model interpretability analysis.
• To provide actionable decision support for banks to prevent customer churn and improve retention.

Main Methods:

• Utilized bank customer data from the Kaggle platform for analysis.
• Employed multiple sampling methods, including SMOTEENN, SMOTE, and ADASYN, for data balancing.
• Developed a customer churn prediction model using a Genetic Algorithm-XGBoost (GA-XGBoost) classifier.
• Conducted interpretability analysis using Shapley values to understand feature impact.

Main Results:

• SMOTEENN demonstrated superior performance in balancing imbalanced banking data compared to SMOTE and ADASYN.
• The GA-XGBoost model achieved optimal F1 and AUC scores of 90% and 99%, respectively, outperforming six other machine learning models.
• Key churn predictors identified include total transactions, transaction amounts, number of products owned, and total sales balance.

Conclusions:

• The GA-XGBoost model provides an effective solution for identifying and predicting customer churn in the banking sector.
• Interpretable insights from Shapley values enable banks to understand churn drivers and enhance customer service quality.
• The study offers valuable references for commercial banks and other industries aiming to reduce customer churn and mitigate losses.