Multiple instance learning for credit risk assessment with transaction data

As the number of personal loan applications grows rapidly, credit risk assessment has become increasingly crucial to both practitioners and researchers. In a traditional assessment system, individual socio-demographic information and loan application information are designed as input for feature engineering; however, an applicant's dynamic transaction history, which is in fact an important indicator for the applicant's pay back behavior, is not included. The present study proposes a comprehensive assessment method that incorporates both conventional data, such as individual socio-demographic information and loan application information, and data for the applicant's dynamic transaction behavior. Our method is based on Radial Basis Function (RBF) Multiple Instance Learning (MIL), which extracts features from a person's transaction behavior history. Five real-world datasets from two large commercial banks in China are used to validate the effectiveness of our proposed method. The experimental results show that our method remarkably improves the prediction performance by using the most commonly used model evaluation criteria.