Counterfactual-Based Synthetic Case Generation

Anik Sen, Mallika Mainali, Christopher B. Rauch, Ursula Addison, Michael W. Floyd, Prateek Goel, Justin Karneeb, Ray Kulhanek, Othalia Larue, David H. Ménager, Matthew Molineaux, J. T. Turner, Rosina O. Weber

Abstract: Case augmentation is often desirable when applying case-based reasoning to real-world problems. Initially explored for explainability, counterfactuals were recently recommended as a strategy to augment data. In this work, we implement an existing approach for generating counterfactuals, propose one variant of the original approach, and propose a third approach based on the literature on algorithmic recourse. We apply these three approaches to two datasets in military medical triage. To assess generalization, we also examine one of our approaches on three publicly available datasets. We compare the approaches based on the number of counterfactuals they produce, their resulting accuracy, overlapping counterfactuals, and domain knowledge. Experimental results are encouraging for the proposed approaches and bring up opportunities for future research.