Analysing the Impact of Data Distribution Shifts on Model Fairness in Machine Learning

As decision-making processes are increasingly automated by the deployment of machine learning techniques, being able to a priori ensure the fairness of these models has become a concern of paramount importance. This is especially true in high-stake domains, where data are often provided as they are, i.e., without any additional insight from domain experts about the presence of possible sensitive attributes, such as gender, nationality, or religion. At the same time, data distribution may evolve over time, i.e., lead to drifts, potentially harmful for the performance of the deployed models. Thus, the interplay between: (i) proactively monitoring for degradation in accuracy and promptly retraining the models, as well as (ii) being able to grant fairness regardless of the possible bias within the data has further entangled this already tricky challenge. In this paper, we present and analyse a synthetically generated example of data distribution shift affecting the model performance, including its fairness. Then, we show how only focusing on singularly addressing either the accuracy drop, rather than the introduced bias, cannot completely solve the issues. In conclusion, we hint at the need for a holistic approach to mitigate both the problems as a possible research direction in this field.