Machine learning models are increasingly used to allocate resources and opportunities like job interviews or kidney transplants. While these models strive to be fair, researchers argue that fairness methods alone are insufficient. They propose a novel approach: randomizing model decisions in a structured way.
Existing Fairness Methods Fall Short:
- Adjust features the model uses for decision-making.
- Calibrate model scores.
- Limitations:
- Don’t address structural injustices and inherent uncertainties.
- Can amplify existing inequalities.
Randomized Decisions Offer Advantages:
- Prevents one individual or group from being consistently disadvantaged.
- Especially beneficial when:
- Model decisions involve uncertainty.
- The same group consistently receives negative decisions.
- Example: Multiple companies using the same deterministic model to rank candidates could always rank the same deserving individual last.
Introducing Structured Randomization:
- Weighted lottery: Allocates resources based on varying levels of randomization.
- Tailored to specific situations: Considers uncertainty and individual claims.
- Improves fairness without compromising model accuracy or efficiency.
Example: Kidney allocation based on projected lifespan can be uncertain. Randomizing decisions in such cases can lead to fairer outcomes for individuals without significantly affecting utility.
Randomization Not a Universal Solution:
- May not be appropriate in contexts like criminal justice.
- Researchers plan to study other use cases like college admissions.
Key Takeaways:
- Randomization offers a novel approach to improve fairness in AI decision-making.
- Provides a tool for stakeholders to consider when allocating scarce resources.
- Further research is needed to explore its applications and potential impact.