Graph-Fused Multivariate Regression via Total Variation Regularization

TL;DR

This paper introduces GFMR, a novel total variation regularized multivariate regression method, demonstrating its efficiency, scalability, and superior performance through simulations and real neuroimaging and activity tracking data.

Contribution

The paper presents a new scalable algorithm for GFMR with proven convergence and statistical consistency, applicable to high-dimensional neuroimaging and physical activity data.

Findings

GFMR outperforms existing methods in simulations

The algorithm is efficient and scalable in distributed platforms

Demonstrated applicability on neuroimaging and activity tracking data

Abstract

In this paper, we propose the Graph-Fused Multivariate Regression (GFMR) via Total Variation regularization, a novel method for estimating the association between a one-dimensional or multidimensional array outcome and scalar predictors. While we were motivated by data from neuroimaging and physical activity tracking, the methodology is designed and presented in a generalizable format and is applicable to many other areas of scientific research. The estimator is the solution of a penalized regression problem where the objective is the sum of square error plus a total variation (TV) regularization on the predicted mean across all subjects. We propose an algorithm for parameter estimation, which is efficient and scalable in a distributed computing platform. Proof of the algorithm convergence is provided, and the statistical consistency of the estimator is presented via an oracle inequality. We present 1D and 2D simulation results and demonstrate that GFMR outperforms existing methods in most cases. We also demonstrate the general applicability of the method by two real data examples, including the analysis of the 1D accelerometry subsample of a large community-based study for mood disorders and the analysis of the 3D MRI data from the attention-deficient/hyperactive deficient (ADHD) 200 consortium.