Modeling Selection for Credit Risk Measurement: Based on Meta Path Features

As the application scope of machine learning expands, studies of credit risk measurement have also witnessed extensive development. An increasing number of studies showed that models based on machine learning algorithms could be used as a substantive solution for credit risk modeling. Recently, path-based features showed their advantages in risk measurement of the rich semantic and relational information it contains. However, studies have yet to be probed into the field that combines meta path features and machine learning models to measure the credit risk of the enterprise. In response to this problem, we compare the performance of machine learning models in terms of meta path features to find suitable machine learning models for meta path features. This paper compares six commonly used machine learning classification models, including neural networks, support vector machine, k-nearest neighbor, random forest, AdaBoost, and GBDT. Experiments on three listed small and medium-sized enterprises datasets in China. We found that neural networks, support vector machines, and the GBDT model perform better than other machine learning models. These are potential classifiers for small and medium-sized enterprises' credit risk measurement.