The R package psvmSDR: A Unified Algorithm for Sufficient Dimension Reduction via Principal Machines

TL;DR

The paper introduces the psvmSDR R package that implements a unified, efficient algorithm for both linear and nonlinear sufficient dimension reduction, including real-time updates, based on principal machine estimators.

Contribution

It presents a new R package implementing principal machine estimators for SDR, extending capabilities to nonlinear and real-time scenarios.

Findings

Efficient computation of SDR estimators using descent algorithm.

Supports both linear and nonlinear SDR methods.

Provides real-time update functionality.

Abstract

Sufficient dimension reduction (SDR), which seeks a lower-dimensional subspace of the predictors containing regression or classification information has been popular in a machine learning community. In this work, we present a new R software package psvmSDR that implements a new class of SDR estimators, which we call the principal machine (PM) generalized from the principal support vector machine (PSVM). The package covers both linear and nonlinear SDR and provides a function applicable to realtime update scenarios. The package implements the descent algorithm for the PMs to efficiently compute the SDR estimators in various situations. This easy-to-use package will be an attractive alternative to the dr R package that implements classical SDR methods.