In Short:
Organizations are using machine-learning models to make decisions like screening job candidates or ranking transplant patients. However, current fairness methods may not address structural injustices. Researchers suggest that randomizing decisions can improve fairness in certain situations, preventing bias. They propose a framework to introduce randomization based on uncertainty levels. Randomization could lead to fairer outcomes without significantly affecting model utility. The researchers plan to study other potential applications of randomization in the future.
Organizations are increasingly using machine-learning models to make decisions around resource allocation and opportunities. These models assist in processes like screening job candidates and prioritizing patients for transplants based on their survival chances.
Addressing Fairness in Machine Learning Models
When implementing a model, users aim to ensure fairness by reducing bias through various techniques. However, researchers from MIT and Northeastern University suggest that current fairness methods may not be enough to tackle structural injustices and uncertainties. They propose a structured randomization approach to enhance fairness in certain scenarios.
The researchers highlight that randomization can prevent deserving individuals from consistently missing out on opportunities. By introducing controlled randomization into model decisions, fairness can be improved without compromising efficiency or accuracy.
Shomik Jain, a graduate student at Institute for Data, Systems, and Society (IDSS), emphasizes the importance of considering uncertainties and fairness in allocating scarce resources using algorithms.
Research Findings
The researchers build on a previous study that revealed deterministic systems can reinforce biases and inequalities, especially when scaling operations. By incorporating randomization strategically, fairness and individual claims can be better respected.
The concept of randomization to improve fairness was found to be valuable from both systemic and individual perspectives, according to Ashia Wilson, the senior author of the paper.
Randomization is seen as a means to uphold the diverse claims individuals have on scarce resources. For example, in kidney allocation, fairness should consider all claims and not favor those with stronger claims.
Benefits of Structured Randomization
The researchers advocate for a weighted lottery system to tailor the level of randomization based on the uncertainty present in decision-making. This method can enhance fairness without compromising the model’s utility.
By calibrating randomization based on statistical uncertainty, fair outcomes can be achieved without significantly impacting the model’s effectiveness.
While randomization may not be suitable for all scenarios, it holds promise in areas like college admissions. Future studies will explore how randomization can improve fairness in other domains and enhance the robustness of machine-learning models.
Ashia Wilson emphasizes that the research serves as a starting point to demonstrate the benefits of randomization in decision-making processes, giving stakeholders a tool to enhance fairness based on their specific requirements.
