Development and Discussion of Distributed Island PSO-based Feature Filtering and GPT model for Credit Risk Assessment

抄録

Credit risk assessment (CRA) never fails to be a common and vital task for financial institutions. To prevent large-scale defaults and even financial turmoil and crisis events, the CRA process contributes to reducing such losses and risks at an early stage. This paper proposes a two-stage calculation framework for this CRA task. Firstly, we developed a novel heuristic algorithm called distributed island particle swarm optimization (DIPSO), extracting the critical features of credit risk data. Then, to obtain the assessment results, we discussed the applicability of the GPT2 model comparing with typical machine learning classifiers. In the experiments, two open-sourced datasets, which are Taiwan Credit Card Dataset and China UnionPay Dataset, are utilized to evaluate the proposed method. The results show that the proposed DIPSO could effectively advance the detection ability for default samples, and the GPT2 model seems to be efficient for raw dataset while failing to perform well on augmented ones. The best solution for CRA should be the combination of ADASYN-based data augmentation and DIPSO with SVM classifier, which achieves high Recall values of 0.6679 and 0.7171 on the two datasets.