Evaluating AI fairness in credit scoring with the BRIO tool

This paper introduces BRIO, a model-agnostic tool for assessing fairness in AI credit scoring systems, and applies it to the German Credit Dataset to analyze and mitigate biases related to sensitive attributes, ultimately aiming to ensure ethical AI practices in financial services.

Abstract

We present a method for quantitative, in-depth analyses of fairness issues in AI systems with an application to credit scoring. To this aim we use BRIO, a tool for the evaluation of AI systems with respect to social unfairness and, more in general, ethically undesirable behaviours. It features a model-agnostic bias detection module, presented in \cite{DBLP:conf/beware/CoragliaDGGPPQ23}, to which a full-fledged unfairness risk evaluation module is added. As a case study, we focus on the context of credit scoring, analysing the UCI German Credit Dataset \cite{misc_statlog_(german_credit_data)_144}. We apply the BRIO fairness metrics to several, socially sensitive attributes featured in the German Credit Dataset, quantifying fairness across various demographic segments, with the aim of identifying potential sources of bias and discrimination in a credit scoring model. We conclude by combining our results with a revenue analysis.

Method

The authors used a methodology centered around the BRIO tool, which comprises a bias detection module and an unfairness risk evaluation module. They applied this tool to the UCI German Credit Dataset to analyze socially sensitive attributes such as gender, age, and nationality. Additionally, they constructed a machine learning model using the Optibinning Python library for credit score prediction and evaluated its performance with metrics like the area under the ROC curve (AUC) and the Gini index. The fairness violation analysis within BRIO employs divergence measures like the Kullback-Leibler and Jensen-Shannon divergences to quantify differences in probability distributions between ideal behavior or different sensitive classes. The risk assessment module in BRIO aggregates the results of fairness tests to provide an overall risk measure of the AI system's unfair behavior.

Main Finding

The authors discovered that the machine learning model constructed using the Optibinning Python library for credit score prediction demonstrated good discriminatory capability when evaluated with metrics such as the AUC and Gini index. Upon applying the BRIO tool to the UCI German Credit Dataset, they found that the model did not significantly amplify biases present in the data related to the sensitive attributes analyzed. This suggests that the model is generally fair in its treatment of different demographic segments. Additionally, the study explored the relationship between fairness risk management and revenue generation, indicating that adjusting the acceptance threshold for credit scores could help balance fairness and profit.

Conclusion

The conclusion of the research is that the BRIO tool is effective in evaluating and ensuring fairness in AI credit scoring systems. The study, which focused on the UCI German Credit Dataset, found that the machine learning model developed did not introduce substantial additional bias beyond what was already present in the dataset. This suggests that the model is fair in its treatment of different demographic segments based on sensitive attributes such as gender, age, and nationality. The research also highlights the potential for balancing fairness and profit in credit scoring practices by managing fairness risks and adjusting credit score acceptance thresholds. The authors suggest that future work will extend these methods to other datasets and contexts, refining the tools and techniques used to evaluate and mitigate bias in AI-driven credit scoring systems.

Keywords

Bias Detection, Unfairness Risk Evaluation, BRIO Tool, Machine Learning, Optibinning, Python, UCI German Credit Dataset, Sensitive Attributes, Demographic Segments, Bias Mitigation, Discrimination, Ethical Guidelines, Revenue Analysis, Kullback-Leibler Divergence, Jensen-Shannon Divergence, ROC Curve, Gini Index, Predictive Accuracy, Decision-Making Processes, Financial Industry, Trust, Social Justice, Equality, Disparate Impact Analysis, Demographic Parity, Equal Opportunity Criteria, Trade-Off, Predictive Modeling, Binning Process, Scorecard, Default Probability, Sensitive Classes, Model Performance, Fairness Metrics, Biases, Algorithmic Interventions, Financial Inclusion, Discriminatory Power, Calibration Measures, Stakeholders, Fairness Violation Analysis, Risk Measurement, Probability Distributions, Threshold Setting, Bad Rate, Provisions, Revenue Generation, Ethical AI, Explainability, Logic Programming, Trustworthiness, Probabilistic Computations, Typed Natural Deduction System, Big Data