Talk 2 on Human-AI Cooperation for Fairness Elicitation
Talk, Fair and Explainable Decision Making (FED) lab, University of Massachusetts Amherst, Amherst, Massachusetts
The talk delves into the intricate concept of fairness and the challenges it poses when attempting to mathematically quantify it for decision-makers. We explore techniques for extracting individuals’ fairness concepts through straightforward inquiries and comparisons. To express fairness concepts mathematically, we employ the power mean function family $\text{M}$$p$$(s; w)$, where $p$ represents the fairness concept, and $s$ and $w$ represent the utility vector and probability measure for the groups. The introduction of the supremum distance metric $\Delta$ allows us to assess the maximum disparity between concepts in various scenarios. Recognizing the computational complexity of $\Delta$ due to the supremum operation, we introduce the additive supremum distance bound function $\Delta_{\uparrow}$ which provides upper bounds for the supremum distance metric. Additionally, we present two practical and efficient supremum distance bound functions, which are proportional to the harmonic difference and log ratio of any pair of fairness concepts ($p, p’$). Furthermore, we demonstrate how to leverage and modify the bounded/unbounded binary search to effectively identify human cardinal fairness concepts.