EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2693. A distributionally robust optimisation approach to fair credit scoring

Invited abstract in session TC-31: Analytics and the link with stochastic dynamics III, stream Analytics.

Tuesday, 12:30-14:00
Room: 046 (building: 208)

Authors (first author is the speaker)

1. Pablo Casas
Decision analytics and risk, University of Southampton
2. Huan Yu
University of southampton
3. Christophe Mues
Southampton Business School, University of Southampton

Abstract

Optimization under uncertainty has been a topic of large debate in Operations research (OR) and Machine Learning (ML) communities for a long time. Recently, many authors have turned to distributional robust optimization (DRO) in search of a solution; however, fewer studies have focused on enhancing fairness under uncertainty. This is of great interest in those applications of ML that can have a direct damaging impact on the population, credit scoring (CS) being one of the most potentially damaging fields according to regulatory bodies.



The paper presented explores the effects of using a DRO-based logistic regression (LR), one of the most commonly used classifiers in CS, across multiple CS datasets. This study will show how robustness has a greater impact on fairness than the fairness constraint, and that the impact on performance is negligible and, in some datasets, positive. We will also provide an empirical analysis of the effect of the different hyperparameters that are unique to DRO-based LR. Furthermore, we will argue that the level of robustness narrows the dispersion of the probability of default distribution and that the parameters in charge of the ground metric have an unnoticeable impact. On a side note, we suggest traditional fairness metrics used in credit scoring are not best suited for the task.

Keywords

Status: accepted

Back to the list of papers