Mitigación de sesgo de género en un modelo de calificación crediticia
Fecha
2024-02-16
Tipo
tesis de maestría
Autores
Corrales Barquero, Ricardo
Título de la revista
ISSN de la revista
Título del volumen
Editor
Resumen
El presente documento expone un trabajo realizado sobre un conjunto de datos y un modelo matemático para apoyo de toma de decisiones en el proceso de crédito para clientes ya constituidos en un banco comercial de Costa Rica. El objetivo principal consistió en evaluar alternativas para mitigar los sesgos de género presentes en el modelo. Para lograrlo, se comenzó por identificar posibles fuentes de sesgo en el modelo, entre las cuales se identificaron posibles sesgos de tratamiento dispar, asociación, selección, sesgo malicioso y sesgo de automatización. Seguidamente se midieron dichos sesgos en más detalle, encontrando que son pequeños, excepto quizá por el sesgo de selección. En tercer lugar, se construyeron modelos alternativos que mitigaran estos sesgos, para finalmente, evaluar la diferencia tanto en las medidas de justicia que se utilizaron como en el rendimiento de los modelos alternativos respecto al original para determinar el que provee mayor valor al negocio. Aquí se encontró que las ganancias son menores y que lo que podría valer más la pena es mantener el modelo actual e investigar otros modelos de calificación crediticia utilizados en otras etapas del proceso de otorgamiento de crédito.
This document presents a project carried out on a dataset and a mathematical model to support decision making in the credit process for established clients in a commercial bank in Costa Rica. The main objective was to evaluate alternatives to mitigate the gender biases present in the model. To achieve this, possible sources of bias in the model were identified, among which possible disparate treatment, association, selection, malicious, and automation biases were identified. These biases were then measured in more detail, finding that they are small, except perhaps for the selection bias. Thirdly, alternative models were built to mitigate these biases, to finally evaluate the difference both in the fairness measures that were used and in the performance of the alternative models compared to the original to determine the one that provides greater value to the business. Here, it was found that the gains are minor and that what could be more worthwhile is to maintain the current model and investigate other credit scoring models used in other stages of the credit granting process.
This document presents a project carried out on a dataset and a mathematical model to support decision making in the credit process for established clients in a commercial bank in Costa Rica. The main objective was to evaluate alternatives to mitigate the gender biases present in the model. To achieve this, possible sources of bias in the model were identified, among which possible disparate treatment, association, selection, malicious, and automation biases were identified. These biases were then measured in more detail, finding that they are small, except perhaps for the selection bias. Thirdly, alternative models were built to mitigate these biases, to finally evaluate the difference both in the fairness measures that were used and in the performance of the alternative models compared to the original to determine the one that provides greater value to the business. Here, it was found that the gains are minor and that what could be more worthwhile is to maintain the current model and investigate other credit scoring models used in other stages of the credit granting process.
Descripción
Palabras clave
ŽŒCREDIT RATING, GENDER, BIAS, ŽŒARTIFICIAL INTELLIGENCE