Regulating Data Bias in Intelligent Decision Making

Artificial Intelligence and Machine Learning can facilitate criminal and juvenile justice systems by augmenting decision-making processes with data-driven insights. However, conventional machine learning techniques may be suffering from unfair predictions induced by historical biases inherent in the training data. In this proposal, we pursue systemic solutions to regulate the historically biased training data. By fitting machine learning models with the bias-regulated training data, we can expect fair and impartial predictions from any regular pipelines and models. The project consists of two research tasks: fixing existing biased data via sample reweighing, and augmenting biased data with generative data retrieval. The first task aims to identify data that may induce unfairness in outcomes, and adjust their weights accordingly in the model's optimization. The second task aims to augment the dataset to make it diverse and representative. It firstly obtains data via generative models that could benefit the fairness in predictions, and then retrieve similar data in external data resources. The entire research plan is illustrated with formulations, technical details, preliminary experimental results, evaluation plans, as well as ways to disseminate the research products to criminal justice practitioners and the broader public. The proposal also discusses the capabilities and competencies of the principal investigator (PI) of this project. With the previous research experience in machine learning fairness, the academic environment, and the supporting resources, the PI has high confidence in successfully implementing the proposed techniques. This Graduate Research Fellowship will serve as a critical stepping stone in the PI's career path in advancing fair and impartial intelligent decision-making, and its applications to criminal and juvenile justice. CA/NCF