Year: 2025
Type: article
Abstract: Financial credit scoring determines whether loan applications are accepted or denied. Missing-not-at-random selection bias results from the fact that we are only able to witness default/non-default labels for authorised samples and are unable to view rejected samples. Such biassed data will always lead to faulty machine learning models. Based on both theoretical analysis and real-world data investigation, we discover a strong correlation between the default/non-default classification task and the rejection/approval classification problem in this work. Therefore, rejection/approval may help with default/non-default learning. As a result, we initially suggest using Multi-Task Learning (MTL) to model the skewed credit score data. In particular, we present a unique Reject-aware Multi-Task Network (RMT-Net) that uses a gating network based on rejection probability to learn the task weights that govern the information passing from the rejection/approval task to the default/non-default task. RMT-Net makes use of the relationship between the two tasks, which states that the default/non-default task must learn more from the rejection/approval task the higher the rejection probability. Additionally, in order to simulate situations with different rejection/approval techniques, we expand RMT-Net to RMT-Net++. Numerous datasets are the subject of extensive testing, which firmly confirms RMT-Net's efficacy on both accepted and rejected samples. Furthermore, RMT-Net++ enhances RMT-Net's functionality.
Authors: Shaik Shafath, B Purushotham
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: diamond
OpenAlex ID: W4410001011