FAIRNESS IN MACHINE LEARNING AND OPTIMIZATION: APPLICATIONS IN COMPUTATIONAL CRIMINOLOGY

Abstract

This research leverages machine learning, optimization, and systems modeling to address fairness in treatment and criminal justice. This dissertation offers an overview of fairness in machine learning, discussing how bias can be introduced through data and model selection, and how fairness metrics such as demographic parity, disparate impact, equalized odds, and statistical parity difference can be employed to identify and mitigate inequality. In addition, the dissertation generalizes these concepts to optimization, pre-processing methods such as reweighting methods, and social welfare metrics such as the Gini coefficient, the Hoover index, which can be embedded into optimization functions. The facility location problem is solved to increase access to drug and alcohol rehabilitation for Oklahoma facilities. Through the lens of distance measurements such as Haversine, Euclidean, Manhattan, and Chebyshev, and constraint conditions such as fairness, the model seeks to find placement of facilities that can minimize travel cost and achieve better fairness for service delivery. Contrary to the common assumption that equity is expensive, adding fairness constraints reduced total travel cost while improving access. Simply, fairness pulls facilities slightly away from already well served corridors toward under served counties, which cuts many long trips and improves overall coverage. We also find that this equity shift pairs naturally with a grid like distance, reinforcing both lower cost and fairer access. Next, compartmental system models are fitted to explore the consequences of treatment-oriented and punishment-oriented strategies to incarceration and recovery. Results suggest that treatment prioritization may lower incarceration rates and costs. Extending this more, an epidemic model where income governs the addiction levels (SIRV2 model) is considered. The model demonstrates that poverty influences outcomes, influences relapse patterns and constrains treatment options, and resource distribution based on such considerations of fairness enhances treatment equity. Lastly, predictive machine learning (ML) models are constructed with the Substance Abuse and Mental Health Services Administration (SAMHSA) rehabilitation's data set in predicting treatment completion. Various methods are evaluated — support vector machines (SVM), decision trees, random forest, and neural networks — as well as fairness interventions like over-sampling, over-sampling using intersectionality, and over-sampling to the worst-case ratio. The results indicate that decision trees and random forests strike the best trade-off between accuracy and fairness. In summary, this work connects algorithmic fairness, optimization, and system modeling with policy that can be made in practice. It offers policy recommendations to improve rehabilitation, decrease incarceration, and foster more equitable systems of addiction recovery.