Co-teaching strategies for credit risk assessment in the presence of label noise

Alfonso Galván Fernández (Student)

Student thesis: Master's Thesis

Abstract

This paper investigates the effect of co-teaching strategies on the handling of reject inference. The investigation explores the utility of co-teaching, a technique known for effectively handling noisy labels, specifically in computer vision, when applied to tabular data for the purpose of understanding its value in credit scoring models. This research provides a comprehensive analysis by comparing two different co-teaching strategies, fixed and incremental forget rates, to traditional the neural network approach in addressing reject inference issues. The findings suggest that although co-teaching has proven effective in other domains, the implementations that this study proposes for credit risk assessment do not provide a significant benefit in accurately predicting defaulting applications. The results indicate that traditional neural networks exhibit greater success in handling predictions for the minority class, which is crucial from a business standpoint. For future research, the study highlights the significance of improving co-teaching strategies, such as modifying the forget rate and investigating alternative network architectures. Future studies can further improve noise-robust learning techniques and their application in complex real-world scenarios by expanding current methodologies and exploring new adaptations. The research contributes to the existing literature by exploring the potential of co-teaching and providing a comprehensive analysis of its effectiveness in the presence of noisy labels. The study provides a basis for future investigations and practical implementations, with the goal of expanding the use and effectiveness of co-teaching strategies in the field of credit risk modeling.