Towards XAI for Optimal Transport
Transport phenomena (or distribution shifts) arise in many disciplines and are often of great scientific interest. Machine learning (ML) is increasingly used in conjunction with optimal transport (OT) to learn models for these. While XAI has improved the transparency of ML models, there has been little discussion on how to explain the factors that drive a distribution shift. Specifically, the issue of opening the OT black box has only received limited attention. Traditional classification models can distinguish between two distributions, but post-hoc explanations based on their gradients may not reveal the true reasons behind their differences. Our goal is to make OT explainable and establish XAI-OT to generate more accurate explanations for distribution shifts. We also discuss concerns regarding the accuracy of optimal transport in the presence of data issues, which we assume to have implications beyond explanations.
