Conformal Prediction for Dyadic Regression Under Complex Missingness

Researchers have proposed new frameworks for conformal prediction and imbalanced classification, addressing complex missingness mechanisms and capacity constraints in machine learning models.

A framework for conformal prediction in dyadic regression problems under complex missingness mechanisms has been developed, establishing asymptotic validity under a nonparametric graphon model^[1]. Additionally, a method called Audited Conformal Prediction (ACP) is proposed for uncertainty quantification in classification models under unknown distribution shift, leveraging a small labeled dataset to train an auxiliary audit model^[2]. ACP produces prediction sets that guarantee marginal coverage and achieves substantially higher conditional coverage in practice than existing approaches. Furthermore, a new framework for imbalanced classification under capacity constraints is proposed, maximizing sensitivity on the minority class while accounting for real operational constraints^[3]. This framework is implemented on top of standard learning methods, including k-NN, SVM, random forests, and neural networks, with established statistical consistency for each method. Experiments on synthetic and real-world datasets validate the proposed methods, with the capacity-constrained classifiers substantially outperforming classical approaches under high imbalance regimes.