Bridging Mechanism Design and Machine Learning towards Algorithmic Fairness
Jessica Finocchiaro and Faidra Monachou wanted to share their experiences while working on the project Bridging Mechanism Design and Machine Learning towards Algorithmic Fairness that got published at ACM Conference on Fairness, Accountability, and Transparency 2021.
In March 2020, the MD4SG working group on Fairness, Bias, and Discrimination got together to launch a series of research projects. Besides ourselves, Roland Maio, Gourab Patro, Manish Raghavan, Ana-Andreea Stoica, and Stratis Tsirtsis were interested in Equitable Mechanism Design. We were from different backgrounds not only by education but also from different parts of the World. In our first meetings, we quickly realized that while the mechanism design (MD) and machine learning (ML) communities both think about fairness, it is often modeled in different ways. One of the most interesting parts about working on this project was learning from collaborators who sit in different departments and fields. Often, we get very comfortable thinking about research with the lens we are calibrated to and comfortable with; working across institutions and global viewpoints empowered us to learn from each other to bridge some of the language and lessons across the fields of machine learning and mechanism design.
After a few meetings of discussions, the Harvard CRCS AI for Social Good workshop (2020) put out a call for papers that included position papers, and this motivated us to start writing down what we had learned as a group. As we started writing a position paper, we realized how much there was to explore, and we continued to flesh out these ideas, submitting a full version of the paper to FAccT.
This paper attempts to (a) review and contrast meanings of “fairness” in the fields of ML and MD, (b) enumerate some of the lessons that have been learned across fields as well as some future lessons we anticipate being applied, and (c) consider examples where ML and MD can be and have been bridged across a handful of domains.