Fr\'echet Sufficient Dimension Reduction for Random Objects

TL;DR

This paper introduces new methods for dimension reduction involving complex random objects and high-dimensional predictors, extending linear and nonlinear approaches with theoretical guarantees and practical applications.

Contribution

The paper proposes a weighted inverse regression ensemble method and a kernel-based operator for Fréchet sufficient dimension reduction, with theoretical analysis and real data application.

Findings

Method performs well in simulations

Theoretical guarantees established

Effective in handwritten digits analysis

Abstract

We in this paper consider Fr\'echet sufficient dimension reduction with responses being complex random objects in a metric space and high dimension Euclidean predictors. We propose a novel approach called weighted inverse regression ensemble method for linear Fr\'echet sufficient dimension reduction. The method is further generalized as a new operator defined on reproducing kernel Hilbert spaces for nonlinear Fr\'echet sufficient dimension reduction. We provide theoretical guarantees for the new method via asymptotic analysis. Intensive simulation studies verify the performance of our proposals. And we apply our methods to analyze the handwritten digits data to demonstrate its use in real applications.