Reproduce “Provably Minimum-Length Conformal Prediction Sets for Ordinal Classification” on OpenResearch

About this paper

Zijian Zhang, Xinyu Chen, Yuanjie Shi, Liyuan Lillian Ma, Zifan Xu, Yan Yan

Researchers at Washington State University developed min-CPS and min-RCPS, new conformal prediction methods that generate provably minimum-length prediction sets for ordinal classification. These model-agnostic and distribution-free approaches consistently reduce prediction set sizes by 14-15% on average compared to prior methods while guaranteeing marginal coverage.

Key insights

• Ordinal conformal prediction can be formally established as an instance-level minimum-length covering problem, enabling the construction of provably optimal (shortest) prediction intervals.
• The empirical coverage rate exhibits a radial monotonicity property with respect to the calibration threshold, allowing for efficient determination of the optimal threshold using binary search.
• Incorporating a length regularization term into the cumulative probability constraint of the prediction set construction can effectively reduce interval sizes further without compromising marginal coverage validity.

Summary by alphaXiv.