Banca de DEFESA: VINICIUS DE OLIVEIRA WATANABE

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : VINICIUS DE OLIVEIRA WATANABE
DATE: 04/03/2026
TIME: 14:30
LOCAL: A defesa será realizada por videoconferência
TITLE:

Fairness Score: A Composite Metric for Fairness and Performance Evaluation

KEY WORDS:

algorithmic fairness; credit risk; fairness metrics; imbalanced data; machine learning; simulation; German Credit.

PAGES: 50
BIG AREA: Ciências Sociais Aplicadas
AREA: Economia
SUMMARY:

The rapid adoption of machine learning in critical domains has intensified concerns about algorithmic bias and the fairness-performance trade-off. While existing literature often focuses on mitigation techniques, model evaluation still tends to assess performance and fairness in isolation. This study proposes a unified evaluation framework centered on a Fairness Score (FS) on a common [0, 1] scale. The framework builds on the Joint Fairness-Performance Index (JFPI), which combines predictive performance (F1-score) and a fairness composite derived from Demographic Parity Ratio, Equalized Odds Ratio, and Predictive Rate Parity. The framework then applies structural adjustments that penalize extreme imbalance and enforce a minimum justice requirement only when a model violates it. We treat the minimum justice threshold as the exogenous policy input and illustrate it with the Four-Fifths Rule. Given this requirement, the framework calibrates the weighting parameter a endogenously from the evaluated model set, rather than treating it as a discretionary preference. We evaluate the framework through controlled Monte Carlo experiments on synthetic credit-risk datasets and assess external validity on the German Credit benchmark. The empirical analysis shows that high-capacity models, especially neural networks, dominate the Pareto frontier in high-bias regimes and outperform linear baselines. Among mitigation strategies, pre-processing interventions provide the most robust results by improving fairness while preserving model stability and predictive integrity. The proposed framework provides a practical tool for responsible and reliable automated decision-making under explicit institutional fairness requirements.

COMMITTEE MEMBERS:
Presidente - 1642911 - DANIEL OLIVEIRA CAJUEIRO
Interno - 1550794 - JOSE GUILHERME DE LARA RESENDE
Externo ao Programa - 1997721 - HERBERT KIMURA - UnBExterno à Instituição - ROBERT ALDO IQUIAPAZA COAGUILA - UFMG