Density-sorted prediction set: Efficient conformal prediction for multi-target regression

Rui Luo, Zhixin Zhou

Abstract: highlights•Introduces DSPS, a new method for uncertainty quantification in multi-target regression that builds flexible, non-convex prediction regions with guaranteed coverage probabilities.•Integrates Conditional Normalizing Flows (CNFs) with conformal calibration to effectively model and optimize the conditional distribution of multiple response variables, addressing complex and nonlinear dependencies.•Utilizes the conditional probability density (derived from the Jacobian determinant and latent density) within CNFs to identify high-density areas in the original response space, ensuring prediction regions are adaptive and focused on regions with high probability density.•Demonstrates through experiments on synthetic and real-world datasets that DSPS produces smaller, more informative prediction regions while maintaining robust coverage, outperforming existing methods like ST-DQR, NPDQR, and Naïve QR.

External IDs:doi:10.1016/j.patcog.2025.112513