Exploring the impact of a fairness algorithm on a machine learning model in a Bioinformatics Application
Presenter Type
UNO Graduate Student (Masters)
Other
Data Science
Advisor Information
Kathryn Cooper, Ph.D.
Location
CEC RM #201/205/209
Presentation Type
Poster
Poster Size
44 x 48
Start Date
22-3-2024 9:00 AM
End Date
22-3-2024 10:00 AM
Abstract
The use of AI tools is expanding rapidly. These tools are increasingly being adopted in applications where there is potential for AI decisions to impact people in ways related to health and safety, and access to resources or opportunities. Research has shown that biases, whether intentional or unintentional, can impact decisions made by AI models. Reducing or eliminating these biases is a non-trivial task, in part due to complexities surrounding the concept of fairness. It is generally accepted that fairness intervention requires a trade-off in which fairness increases at the expense of reducing accuracy. Additionally, there is no generally accepted framework for if or when fairness intervention is called for, nor is there a standard definition of fairness incorporating all types of harm resulting from AI decisions affected by bias. As a result, caution, awareness, and assessment are necessary when there is potential for AI decisions to impact minority groups. This project explores the use of a specific fairness mitigation tool, Correlation Remover, from the Fairlearn Library, by applying a Random Forest Classifier to a dataset containing observations of diabetes patients. The Random Forest model was used to predict which patients would be readmitted to the hospital within thirty days. Correlation Remover was then used on the data set to remove correlations with the Race category. Random Forest Classifier accuracy and feature importance rankings were measured pre-and-post fairness intervention. Contrary to expectations, use of Correlation Remover had a negligible impact on model accuracy. Additionally, the model’s feature importance rankings did not change significantly.
Exploring the impact of a fairness algorithm on a machine learning model in a Bioinformatics Application
CEC RM #201/205/209
The use of AI tools is expanding rapidly. These tools are increasingly being adopted in applications where there is potential for AI decisions to impact people in ways related to health and safety, and access to resources or opportunities. Research has shown that biases, whether intentional or unintentional, can impact decisions made by AI models. Reducing or eliminating these biases is a non-trivial task, in part due to complexities surrounding the concept of fairness. It is generally accepted that fairness intervention requires a trade-off in which fairness increases at the expense of reducing accuracy. Additionally, there is no generally accepted framework for if or when fairness intervention is called for, nor is there a standard definition of fairness incorporating all types of harm resulting from AI decisions affected by bias. As a result, caution, awareness, and assessment are necessary when there is potential for AI decisions to impact minority groups. This project explores the use of a specific fairness mitigation tool, Correlation Remover, from the Fairlearn Library, by applying a Random Forest Classifier to a dataset containing observations of diabetes patients. The Random Forest model was used to predict which patients would be readmitted to the hospital within thirty days. Correlation Remover was then used on the data set to remove correlations with the Race category. Random Forest Classifier accuracy and feature importance rankings were measured pre-and-post fairness intervention. Contrary to expectations, use of Correlation Remover had a negligible impact on model accuracy. Additionally, the model’s feature importance rankings did not change significantly.