Researchers propose fair budget aggregation for AI systems
New arXiv paper solves fairness in resource allocation algorithms
A new paper addresses fairness in budget aggregation, where individual distributions over alternatives are combined into a collective distribution. The authors define two individual fair share guarantees and show that for utilities derived from ℓ_t metrics with t≥1, both guarantees can be satisfied along with Pareto efficiency, with the distributions computable in polynomial time. For ℓ_1 utilities, they prove that Pareto efficiency, strategyproofness, and a very weak fairness notion called single-minded positive share are not always compatible when n,m ≥ 3, and they provide rules satisfying these three axioms for smaller parameters. They also establish similar impossibility results for ℓ_2 utilities.
- Introduces two individual fair share guarantees for budget aggregation problems in AI systems
- Proves polynomial-time solutions exist for ℓₜ utility metrics (t≥1) that satisfy fairness and efficiency simultaneously
- Demonstrates incompatibility of key axioms (Pareto efficiency, strategyproofness) for ℓ₁ utilities when n,m ≥ 3
Why It Matters
Enables truly fair AI decision-making in resource allocation from hiring to public services