29 August 2024
14:00 Master's Defense Room 85 of IC2
Theme
MAPOFCEM: Mining Feasible Counterfactual Explanations from Pareto-optimal Agnostic Model
Student
Arthur Hendricks Mendes de Oliveira
Advisor / Teacher
Marcos Medeiros Raimundo
Brief summary
Machine learning models are taking an increasing role in credit scoring decision-making due to their accuracy in predicting loan repayment. However, a criticism regarding the implementation of these models is the difficulty of explaining the algorithm's decision-making to individuals whose credit applications have been rejected. Recent studies reveal that counterfactual explanations provide users with model decision feedback through a list of changes they can make to their profile to guide future applications. Therefore, providing viable counterfactual explanations is a crucial factor in ensuring that proposed changes are within the reach of users. We propose a method called ModelAgnostic Pareto-Optimal Feasible Counterfactual Explanations Mining (MAPOFCEM) to provide viable and actionable feedback on decisions made by a credit risk algorithm. This method allows individuals who have been denied a loan to make specific adjustments to their profiles, thereby increasing their chances of loan approval in the future. Our approach integrates an outlier detection mechanism into the counterfactual search process to generate viable counterfactual explanations. The experimental results demonstrate that MAPOFCEM provides a more viable and robust framework compared to existing open source reference models in the literature, which increases the usability of such tools for evaluating credit risk models in real-world applications.
Examination Board
Headlines:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Marcelo da Silva Reis | IC / UNICAMP |
| Luis Gustavo Nonato | ICMC / USP |
Substitutes:
| Saullo Haniell Galvão de Oliveira | PUC-Campinas |
| Leonardo Tomazeli Duarte | FCA/UNICAMP |