Assessing algorithmic fairness without sensitive information

Abstract

As the prevalence of algorithmic decision-making increases, so does the study of algorithmic fairness. When this aspect is disregarded, bias and discrimination are created, reproduced or amplified. Accordingly, work has been done to harmonize definitions of fairness and categorize ways to improve it. While using demographic data about the protected group is a possible solution, in real-world applications privacy concerns as well as uncertainty about the relevant attributes make it unrealistic. Consequently, we seek in this work to provide an overview of the methods that do not require such data, to identify which areas might be under-researched and to propose research questions for the first phase of the PhD. The influence of datasets size in the discovery and mitigation of unknown biases appears to be such an area, one that we plan to explore more fully during the thesis.