Beyond Uncertainty Sets: Leveraging Optimal Transport to Extend Conformal Predictive Distributions to Multivariate Settings
The paper extends Conformal Prediction to multivariate scores by using Optimal Transport (OT). It solves two problems: defining ranks for vector scores and creating finite-sample calibration guarantees for prediction distributions. This method uses OT maps and polyhedral partitions to ensure the resulting prediction sets are tractable and rigorously calibrated, even in high dimensions.