24 April 2024
15:00 Master's Defense IC3 Auditorium
Theme
CI-EX: Confident-Inline Extrapolation for Inference of Rejects in Financial Credit Scores
Student
Athyrson Machado Ribeiro
Advisor / Teacher
Marcos Medeiros Raimundo
Brief summary
One of the main challenges in the area of credit scoring modeling is the unavailability of verifying the real payment capacity of customers who have had their credit proposal denied (rejected customers). More basic credit scoring processes only take into account the population of accepted customers, and may be biased against individuals other than that distribution of individuals. Rejected inference are methods that aim to combat this bias by inferring missing information from rejected individuals and including them in the credit scoring system. Classic reject inference methods are generally based on the assumption that it is possible to extrapolate the behavior of rejected customers from data on accepted customers, which is not always the case. We then propose a semi-supervised method that, at each iteration, identifies and labels only the rejected individuals closest to the distribution of the accepted population. This selection is made through an outlier detection model. Only individuals who show the highest confidence in the predicted labels are incorporated into the new dataset for training. We assumed that labeling rejected individuals closer to the distribution of accepted ones could help combat biases in the inference process. We show that our method is better at identifying positive samples than benchmark reject inference models from the literature, providing a more inclusive and equitable evaluation framework for credit scoring.
Examination Board
Headlines:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Eliezer de Souza da Silva | FGV |
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
Substitutes:
| Anderson de Rezende Rocha | IC / UNICAMP |
| Saullo Haniell Galvão de Oliveira | PUCC |