A new academic framework wants to settle arguments about what "fair" even means in algorithms - by giving every definition a common structure to sit in.
The AR fairness metamodel, described in a new arXiv paper, breaks fairness scenarios into agents, resources, and attributes, then lets researchers plug in different fairness measures for comparison. The paper walks through both discrete and continuous measures - equality, equity, group fairness, individual fairness, the Gini index, the Theil index, and Jain's fairness index - plus a worked example using Australia's Child Care Subsidy. It also proves formal relationships between group fairness, individual fairness, and envy-freeness. The underlying conceptual modeling comes from something called the Tiles framework, whose modular pieces can be recombined to capture different fairness definitions. An open-source implementation of Tiles is available for anyone who wants to try it.
Fairness metrics in machine learning have a well-known problem: there are dozens of them, they often contradict each other, and picking one is usually more of a policy choice dressed up as a technical one. A metamodel that lets you formally compare group fairness against individual fairness, or check how a metric behaves against envy-freeness, is useful less for declaring a winner and more for making the tradeoffs explicit before deployment.
Don't expect this to end the fairness wars. It is a common vocabulary for describing them, which is still progress - most fairness disputes stall out because two teams are arguing past each other with different implicit definitions.