Many software systems today make use of large amount of personal data to make recommendations or decisions that affect our daily lives. These software systems generally operate without guarantees of non-discriminatory practices, as instead often required to human decision-makers, and therefore are attracting increasing scrutiny. Our research is focused on the specific problem of biased software-based decisions caused from biased input data. In this regard, we propose a data labeling framework based on the identification of measurable data characteristics that could lead to downstream discriminating effects. We test the proposed framework on a real dataset, which allowed us to detect risks of discrimination for the case of population groups.

URL: https://www.springer.com/us/book/9783030116798