Resolve Magazine Fall 2023 >> Making Sense of Machine Learning >> Stories >> …and fairer trade-offs
As machine learning is increasingly deployed across a range of sectors—from education and advertising to credit scoring and criminal justice—mitigating bias and improving fairness in these systems has become a critical concern. But achieving both high levels of accuracy and fairness is an elusive goal, as algorithms are typically designed to maximize one or the other.
“In applications like credit scoring, the prediction outcomes might be biased toward underrepresented groups, leading to unfairness,” says Luis Nunes Vicente (pictured), Timothy J. Wilmott Endowed Chair Professor and Department Chair of the Department of Industrial Systems and Engineering. “Achieving perfect fairness might result in reduced accuracy, and vice versa.”
Nunes Vicente and Suyun Liu '22 PhD (pictured below), now an applied scientist at Amazon, designed and implemented multi-objective fair learning algorithms that quantify this trade-off between fairness and accuracy.
“We seek to approximate the entire ‘Pareto front,’ which uniquely defines the trade-offs between conflicting objectives,” says Nunes Vicente. 
“The decision-makers can then themselves use such a trade-off curve to select a solution based on the desired or legally required level of fairness across different demographic groups.”
He says their framework can be applied to systems such as facial recognition and self-driving cars.
“The ultimate goal of this project is to compute the entire Pareto front for fairness-accuracy trade-offs of a given machine learning system,” he says. “In practice, given any constraint on fairness level, decision-makers can leverage the trade-off curve by easily picking the solution with the best model performance.”
Main image: Formatoriginal/Adobe Stock
