Multimodel Integrated Enterprise Credit Evaluation Method Based on Attention Mechanism

Area of Science:

• Financial Technology
• Machine Learning in Finance
• Credit Risk Management

Background:

• Small and medium-sized enterprises (SMEs) face significant challenges in obtaining financing and loans due to difficulties in credit risk assessment.
• Existing methods hinder SME operations and development.

Purpose of the Study:

• To develop a robust and reliable model for assessing credit risk in SMEs.
• To improve the accuracy and efficiency of financing and loan assessments for SMEs.

Main Methods:

• Feature selection using correlation coefficient and Gradient Boosting Decision Tree (GBDT).
• Integration of an attention mechanism (SE-Block) into feature tensors.
• Training XGBoost and LightGBM models on data subsets.
• Fusion of model results using Bayesian ridge regression.

Main Results:

• The proposed NN-ATT-Bayesian-Stacking model achieved an AUC value of 0.9675 in simulation experiments.
• The model demonstrated ideal prediction result distribution and good robustness.
• The model provides a reliable assessment for SME financing and loans.

Conclusions:

• The developed model significantly enhances credit risk assessment for SMEs.
• This approach offers a reliable framework for supporting SME financing and loan accessibility.
• Improved credit risk assessment can foster the growth and development of SMEs.