Multimodel Integrated Enterprise Credit Evaluation Method Based on Attention Mechanism

Abstract

Due to the difficulty of credit risk assessment, the current financing and loan difficulties of small- and medium-sized enterprises (SMEs) are particularly prominent, which hinders the operation and development of enterprises. Based on the previous researches, this paper first screens out features by correlation coefficient method and gradient boosting decision tree (GBDT). Then, with the help of SE-Block, the attention mechanism is added to the feature tensor of the subset separated from metadata. On this foundation, two models, XGBoost and LightGBM, are used to train four subsets, respectively, and Bayesian ridge regression is used to fuse the training results of single models under different subsets. In the simulation experiment, the AUC value of the NN-ATT-Bayesian-Stacking model reaches 0.9675 and the distribution of prediction results is ideal. The model shows good robustness, which could make a reliable assessment for the financing and loans of SMEs.