German Martinez Matilla

German Martinez Matilla, Jakub Marecek

Sample complexity of bias estimation is a lower bound on the runtime of any bias detection method. Many regulatory frameworks require the bias to be tested for all subgroups, whose number grows exponentially with the number of protected attributes. Unless one wishes to run a bias detection with a doubly-exponential run-time, one should like to have polynomial complexity of bias detection for a single subgroup. At the same time, the reference data may be based on surveys, and thus come with non-trivial uncertainty. Here, we reformulate bias detection as a point-to-subspace problem on the space of measures and show that, for supremum norm, it can be subsampled efficiently. In particular, our probabilistically approximately correct (PAC) results are corroborated by tests on well-known instances.

German M. Matilla, Jiri Nemecek, Illia Kryvoviaz et al.

There is a strong recent emphasis on trustworthy AI. In particular, international regulations, such as the AI Act, demand that AI practitioners measure data quality on the input and estimate bias on the output of high-risk AI systems. However, there are many challenges involved, including scalability (MMD) and computability (Wasserstein-1) issues of traditional methods for estimating distances on measure spaces. Here, we present humancompatible.detect, a toolkit for bias detection that addresses these challenges. It incorporates two newly developed methods to detect and evaluate bias: maximum subgroup discrepancy (MSD) and subsampled $\ell_\infty$ distances. It has an easy-to-use API documented with multiple examples. humancompatible.detect is licensed under the Apache License, Version 2.0.