Algorithmic Fairness Datasets: Supporting Principled Data Practices
Experimental results are only as good as their input data.
In machine learning there is a growing awareness that dataset documentation is key to improve data utilization.
Much of this awareness comes from the algorithmic fairness community, through a variety of dedicated frameworks and initiatives.
This begs the question as to whether this scholarly field has benefitted from its own initiatives and attention to data documentation.
In this talk, Alessandro FABRIS, a PhD candidate at the University of Padua, and Gian Antonio SUSTO, currently Associate Professor at the University of Padova, answer in the negative, by describing the current state of data utilization in algorithmic fairness and highlighting some limitations which hinder reliable progress.
To address them, Fabrise will present a tool enabling more principled dataset practices, supporting dataset search from multiple angles, sustained by a wide and standardized documentation effort. Finally, he will present the wider relevance and envisioned applications of this initiative for the research community.