Using ensemble learning and explainable AI to predict bank marketing customer subscription.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Financial Technology

Background:

• Class imbalance and distribution shifts pose significant challenges in bank marketing predictive tasks.
• Existing models often struggle with generalization and interpretability in dynamic financial environments.

Purpose of the Study:

• To develop an advanced predictive model, CBE-XAI (CatBoost Ensemble with eXplainable AI), for term deposit subscription intention.
• To enhance model accuracy, interpretability, and cross-domain generalization in bank marketing.

Main Methods:

• Integration of five heterogeneous CatBoost base learners with dynamic post-training weighting.
• Application of a hierarchical SHapley Additive exPlanations (SHAP) system for feature importance analysis.
• Implementation of a pre-deployment adaptive fine-tuning (AFT) strategy with a composite loss function for model calibration.

Main Results:

• CBE-XAI achieved an AUROC of 0.949 and F1-Score of 0.621 on public datasets, outperforming standard benchmarks.
• The AFT strategy improved AUROC from 0.842 to 0.931 in external validation on a Chinese bank dataset.
• Hierarchical SHAP analysis provided insights into feature importance, aligning with banking business logic.

Conclusions:

• CBE-XAI offers a robust solution for bank marketing by balancing predictive accuracy, interpretability, and generalization.
• The adaptive fine-tuning strategy is crucial for enhancing cross-domain performance.
• Explainable AI methods provide valuable insights into model decision-making processes in financial applications.