Non-negative matrix factorization via projected gradient method for credit risk analysis

Credit risk assessment of financial intermediaries is an essential problem in finance. The key is to find accurate predictors of individual risk in the credit portfolios of institutions. However, accessing credit risk is very difficult because many factors may contribute to the risk and their relationship is complicated to capture. In recent years, machine learning techniques, such as SVM classifier, have been successfully applied into the field of credit risk analysis. SVM is a strong classifier that is effective in capturing nonlinear relationship in the data. However, high dimensional training data not only results in time-consuming computation but also affects the performance of the classifier. In this paper, we will adopt non-negative matrix factorization via project gradient method to transform the data into lower dimensional space that will contribute to good performance in the credit risk classification. We test our method in a real-world credit risk prediction task, and our empirical results demonstrate the advantage of our method by comparing with other state of art methods.