When the Data Change: Will the Model Still Be Fair?

Hristo Inouzhe

Departamento de Matemáticas, Universidad Autónoma de Madrid

Resumen/Abstract

A model may receive a fairness certificate during its design phase and still become unfair once it is put to use. Consider a mortgage-pricing model. It is trained on applications from one year. A year later, interest rates, housing markets, and local economies have changed. Before seeing the new results, can we tell how unequal its predictions might become?This talk offers a practical answer. We start with plausible changes in the population, stated explicitly and calibrated from data. We then obtain a bound on the inequality that the model's predictions could produce. Optimal transport lets us describe realistic shifts rather than arbitrary disturbances.Cases in mortgage pricing, insurance, and other datasets show that comparing group averages alone can hide important differences. The result is a tool for monitoring models and deploying them more safely. 

Curriculum ponente

Hristo Inouzhe is an Associate Professor in the Department of Mathematics at Universidad Autónoma de Madrid. His research develops statistical and machine-learning methods that remain interpretable and reliable when data are functional or high-dimensional, distributions shift, or decisions must be made under uncertainty. His current work combines optimal transport, functional quantiles and depth, robust inference, distribution shift and algorithmic fairness, in-context learning, and applied statistical modelling. He is particularly interested in connecting mathematical guarantees with problems in health, social policy, environmental science, biological modelling, and artificial intelligence.