Home

The goal of this work is the application of logic-based approaches to identify potential proxy attributes (single or combined) in datasets that can lead to discriminatory Machine Learning models. In the identification of proxy attributes, different relations between unprotected and protected attributes are considered, such as equivalence and implication. Moreover, it is considered that a proxy attribute may have a degree of confidence associated. For example, if it is not possible to infer accurately all the values (100%) of a protected attribute from the information of a set of unprotected attributes, but it is possible to infer most of the values (e.g. 90 %), then one may want to consider such set of unprotected attributes as proxy attributes as well.