| With the deep integration of Internet technology and financial industry,deep learning-based credit scoring models gradually replace traditional credit scoring with excellent accuracy to become a hot spot for research.The deep neural network-based credit scoring model,however,is difficult to be widely used in the credit scoring field because of the opaque learning and decision-making process.Therefore,this topic is based on deep learning-based credit scoring models,and studies how to improve the accuracy and interpretability of deep learning-based credit scoring models,respectively,as follows.The current deep learning-based credit scoring models are too dependent on labeled training data and difficult to use rich domain knowledge in credit scoring.In order to integrate deep learning models with credit scoring domain knowledge to improve model classification accuracy,the paper proposes a probabilistic soft logic posterior regularization framework.Firstly,it is proposed to calculate the logical rules with soft logic probability and incorporate the prior knowledge by constraining the posterior distribution of the neural network output to form a logical output that conforms to the rules;secondly,it is proposed to introduce the loss term that measures the true label and logical output into the loss function and use the back propagation of the deep learning training process to effectively convert the logical rules into the weights of the neural network;finally,on the public dataset of credit scoring The experimental results show that the deep credit scoring model based on probabilistic soft logic posterior regularization framework outperforms the traditional deep learning model.The current deep learning-based credit scoring models lack interpretability in decision making,which makes them difficult to be applied in practice.To address these problems,the paper improves Trepan,a decision tree based neural network rule extraction technique,which clarifies the decision making of neural network models by capturing the explanatory rules embedded in the learning knowledge of the neural network models.First,a fidelity-based prepruning method is proposed to make the Trepan decision tree enhance interpretability while ensuring decision accuracy;second,an evaluation system is designed to evaluate the interpretability,accuracy,and fidelity of this type of method completely;finally,the LIME method,a local interpretability technique,is used to analyze the influence of features on decision results during the model decision process.Experiments are conducted on a public dataset of credit scoring,and the experimental results show that the proposed method is able to generate accurate and interpretable credit scoring models.The paper investigates and experimentally validates the deep learning-based credit scoring model in terms of both accuracy and interpretability of the model.The results show that the proposed model has better accuracy and interpretability,which can further promote the application of deep learning models in credit scoring. |
