Tuning parameter selection rules for nuclear norm regularized multivariate linear regression

TL;DR

This paper develops new data-driven rules for selecting tuning parameters in nuclear norm regularized multivariate linear regression, improving model estimation in high-dimensional settings.

Contribution

It introduces four novel tuning parameter selection rules based on duality theory and projection properties, filling a gap in optimization-based parameter tuning for NMLR.

Findings

The rules effectively control the rank of the estimated coefficient matrix.

PSR+ is identified as the most efficient rule.

Numerical experiments demonstrate the rules' practical value.

Abstract

We consider the tuning parameter selection rules for nuclear norm regularized multivariate linear regression (NMLR) in high-dimensional setting. High-dimensional multivariate linear regression is widely used in statistics and machine learning, and regularization technique is commonly applied to deal with the special structures in high-dimensional data. As we know, how to select the tuning parameter is an essential issue for regularization approach and it directly affects the model estimation performance. To the best of our knowledge, there are no rules about the tuning parameter selection for NMLR from the point of view of optimization. In order to establish such rules, we study the duality theory of NMLR. Then, we claim the choice of tuning parameter for NMLR is based on the sample data and the solution of NMLR dual problem, which is a projection on a nonempty, closed and convex set. Moreover, based on the (firm) nonexpansiveness and the idempotence of the projection operator, we build four tuning parameter selection rules PSR, PSRi, PSRfn and PSR+. Furthermore, we give a sequence of tuning parameters and the corresponding intervals for every rule, which states that the rank of the estimation coefficient matrix is no more than a fixed number for the tuning parameter in the given interval. The relationships between these rules are also discussed and PSR+ is the most efficient one to select the tuning parameter. Finally, the numerical results are reported on simulation and real data, which show that these four tuning parameter selection rules are valuable.