A1343
Title: AI-guided single-index varying coefficient logistic modeling: A rain-snow partitioning case study
Authors: Maha Moussa - Utah State University (United States)
Yan Sun - Utah State University (United States) [presenting]
Wei Zhang - Utah State University (United States)
Shandian Zhe - University of Utah (United States)
Abstract: Varying-coefficient (VC) models make effect heterogeneity explicit, but traditional statistical estimation can be computationally demanding and unstable in high-dimensional settings. In parallel, modern artificial intelligence (AI), especially deep learning, delivers highly accurate predictions but typically lacks built-in guarantees of identifiability, shape constraints, or easily interpretable effect summaries. An AI-guided single-index vc logistic framework is developed in which a flexible "teacher" network supplies structured, data-dependent regularization for an interpretable "student" model. The student is a semiparametric single-index VC logistic regression whose parameters are modeled as smooth, shape-constrained functions of a learned contextual index, and are estimated via penalized logistic likelihood with AI-informed penalties. As a case study, the framework is applied to precipitation-phase partitioning (rain-snow classification), using approximately 16.8 million reports from 11,626 stations across the Northern Hemisphere. The resulting estimator attains AI level high test accuracy, while yielding statistically interpretable outputs. More broadly, the framework illustrates how AI-derived structure can stabilize and scale the single-index VC logistic model in complex applications, providing a template for embedding deep-learning targets into penalized semiparametric estimation.
