Faculty of Actuarial Science & Insurance Seminar with Benjamin Avanzi

Abstract:
Ensuring fairness in machine learning models is essential for their application in various fields, particularly in decision-making processes that impact individuals. Recent advancements in machine learning have led to the development of complex models that achieve high predictive accuracy but have also revealed latent societal inequalities in decision-making processes that may have detrimental effects on certain sub-groups in society.

The concept of fairness in machine learning is complex and varies depending on the legal framework and cultural norms. Typically, fairness is mathematically encoded into a set of criteria. Group-level fairness criteria, such as demographic parity, equal opportunity, and equalized odds, emphasize the promotion of equitable treatment among various groups, particularly those identified by protected characteristics like race, gender, or age. On the other hand, individual-level fairness ensures that individuals with similar characteristics receive similar treatment.  A significant portion of the literature on fairness centers around simple binary classification problems, where the protected feature is usually coded as binary. For instance, the widely used COMPAS software in the US, designed
to determine the likelihood of reoffending by inmates, has been thoroughly investigated after being found to be biased against African American individuals. While the focus is primarily on binary classification, there are also cases where regression tasks are affected by a lack of fairness, including insurance pricing. In fact, the European Union Directive 2004/113/EC in conjunction with the Guidelines on the application of Council Directive 2012/C 11/01 prohibits insurance instruments from being priced differently based solely on gender differences. Scenarios can become more complex; e.g., insurance pricing needs to be performed without considering gender differences while avoiding the use of cultural and racial origin. In such cases, it has been observed that even models
constructed to be race- and gender-neutral can still discriminate within subcategories. For example, Black males may receive different treatment than White females, known as intersectional unfairness.
This paper considers data that display latent inequalities and proposes a novel regularization approach based on distance covariance, that aims to reduce their effect on model output while adhering to a pre-specified fairness criterion, such as demographic parity.  The regularizer is compatible with both regression and classification tasks and can be applied to any modeling approach that optimizes an objective. Furthermore, unlike existing methodologies, it allows for the mitigation of multiple protected features of any type, such as categorical or continuous, mitigating the impact
of differences within protected (intersectional) subgroups.

The methodology presented in this study performs competitively with some of the state-of-the-art methods that address a single protected feature. It also offers a ready-to-use statistical test that can be used to calibrate the regularization parameter and validate the satisfaction of the fairness criterion. Moreover, it can be extended to account for multiple protected features jointly, performing well in equal treatment of protected (intersectional) subgroups.

Faculty Events | Website | View More Events

Contact the organizers