Prescriptive modality selection for classification and regression a a b

Emilio Carrizosa

doi:10.1016/j.ejor.2026.07.009

Prescriptive modality selection for classification and regression a a b

2026-08-01

0

PRE

AI

DOI：10.1016/j.ejor.2026.07.009

Original

Original request for help

Save

Abstract

En 中文

This paper presents a prescriptive analytics framework for supervised learning in multimodal settings, where features are organized into blocks (or modalities) with varying acquisition costs. We address the problem of individualized modality selection under budget constraints, determining which information to acquire for each forthcoming individual to improve predictive performance. Our approach has two steps. First, we build a Random Forest obtained by ensembling customized Decision Tree models that explicitly exploit the block-arranged structure of the features, and estimate the prediction errors when different blocks are taken into account. Second, and based on these predictions, modality selection for each forthcoming individual which demands a prescription is formulated as an Integer Programming problem that minimizes an estimate of the overall prediction error subject to budget limits. The proposed framework scales to high-dimensional multimodal settings while offering interpretability by ranking the importance of modality subsets with respect to the response variable. Results on synthetic and real-world datasets show that the method achieves efficient resource allocation for both classification and regression tasks.

Keywords:

Prescription

Mathematical optimization

Supervised learning

Prescriptive modality selection for classification and regression a a b

Abstract

Journal

Organization

Cited Papers