A novel method for credit scoring based on feature transformation and ensemble model

Area of Science:

• Machine Learning
• Financial Analytics
• Data Science

Background:

• Credit scoring is crucial for financial institutions to identify potential defaulters.
• Traditional methods often struggle with data imbalance and manual feature engineering.

Purpose of the Study:

• To propose a novel credit score prediction method.
• To enhance accuracy and efficiency in distinguishing defaulting users.

Main Methods:

• A cascade approach combining feature transformation (Boosting Trees and Auto-Encoders) and a heterogeneous ensemble model (Factorization Machine and Deep Neural Networks).
• Feature transformation addresses data imbalance and replaces manual feature engineering.
• Ensemble model extracts both low-order and high-order feature intersections.

Main Results:

• The proposed method demonstrated superior performance compared to existing credit scoring models.
• Experiments on standard datasets validated the effectiveness of the approach.

Conclusions:

• The developed credit scoring method offers improved accuracy.
• The combination of advanced feature transformation and ensemble modeling provides a robust solution for credit risk assessment.