Algorithmic Fairness in Applied Machine Learning Contexts

Fairness is a common standard in machine learning principles, ethics declarations, and best practices statements. Fairness, though, does not have a singular definition in machine learning. One common theme among fairness concepts is equal treatment by a machine learning model across groups. In some cases, it might be more fair for the machine learning model to produce different predictions or classifications for each group in concordance with differences in their outcome rates. Equal treatment might be an appropriate standard of fairness in some circumstances but not in others. Codes of ethics and standards for machine learning offer many different suggestions about how machine learning ought to be fair. Not only is there a diversity of fairness concepts, but standards also often offer little or no guidance on how these fairness axioms should guide the real-world practice of developing and deploying machine learning models in applied settings. The context in which machine learning is applied may determine which aspects of fairness are expected or upheld. Machine learning to shape, for example, (a) consumer loan approval or rates, (b) job recommendations, (c) text translations, (d) credit decisions, and (e) justice decisions might all impel different conceptions of machine learning fairness. Going beyond an expectation of “equal treatment”, machine learning in each of these areas might think about fairness differently.