Abstract: Black-box machine learning models have gained widespread deployment in decision-making settings across many parts of society, from sentencing decisions to medical diagnostics to loan lending. However, many models were found to be biased against certain demographic groups. Initial work on Algorithmic fairness focused on formalizing statistical measures of fairness, that could be used to train new classifiers. While these models were an important first step towards addressing fairness concerns, there were immediate challenges with them. Causality has recently emerged as a powerful tool to address these shortcomings. Causality can be seen as a model-first approach: starting with the language of structural causal models or potential outcomes, the idea is to frame, then solve questions of algorithmic fairness in this language. Such causal definitions of fairness can have far-reaching impact, especially in high risk domains. Interpretability on the other hand can be viewed as a user-first approach: can the ways in which algorithms work be made more transparent, making it easier for them to align with our societal values on fairness? In this way, Interpretability can sometimes be more actionable than Causality work.

Given these initial successes, this workshop aims to more deeply investigate how open questions in algorithmic fairness can be addressed with Causality and Interpretability. Questions such as: What improvements can causal definitions provide compared to existing statistical definitions of fairness? How can causally grounded methods help develop more robust fairness algorithms in practice? What tools for interpretability are useful for detecting bias and building fair systems? What are good formalizations of interpretability when addressing fairness questions?

Website: www.afciworkshop.org