Fair Engineering of Machine Learning Systems – Lessons Learned from a Literature Review

Abstract

With the growing prevalence of AI algorithms and their use to prepare and even execute decisions, there is increasing debate about whether the results of machine learning systems tend to be fairer or more unfair. When faced with engineering a fair machine learning solution in practice, trade-offs arise between conflicting fairness notions. We conduct a literature review on this topic. The results of our review indicate that a slight consensus exists that the human concept of fairness is much broader than what lies in the scope of current fairness metrics. We discuss the context of judging fairness metrics. We also find that, albeit much research already has been done, there is room for improvement when seeking to generalize the findings across different scenarios.