UC San Diego

Machine learning models have increasingly been used to facilitate the decision-making process in high-stakes tasks. Recourse, referring to the ability to provide actions that the affected individuals can take to achieve a favorable outcome, has received significant attention in research. Previous studies focused on developing algorithms that generate resources with desired properties such as actionability or diversity. However, there is no guarantee that all unfavorable users will receive actionable recourse, and treatment for those no-recourse subjects should also be investigated. This thesis examines three novel approaches of altering the dataset and retraining the model to address this challenge and enhance the actionable recourse available to individuals. This thesis also demonstrates how well-chosen data features can offer an effective solution to this problem through experimentation with both synthetic and real-world datasets.