A re-examination to the SCoTLASS problems for SPCA and two projection-based methods for them

TL;DR

This paper re-examines the SCoTLASS sparse PCA problem, proving solution equivalences, proposing gradient projection and Newton algorithms, and demonstrating their efficiency and scalability through numerical experiments.

Contribution

It introduces new algorithms for SCoTLASS problems based on projection methods, with proven convergence and improved performance over existing methods.

Findings

Solutions to SPCA-P1 and SPCA-P3 are the same.

Solutions to SPCA-P2 and SPCA-P3 are mostly the same.

ANSPCA outperforms GPSPCA on large-scale data.

Abstract

SCoTLASS is the first sparse principal component analysis (SPCA) model which imposes extra l1 norm constraints on the measured variables to obtain sparse loadings. Due to the the difficulty of finding projections on the intersection of an l1 ball/sphere and an l2 ball/sphere, early approaches to solving the SCoTLASS problems were focused on penalty function methods or conditional gradient methods. In this paper, we re-examine the SCoTLASS problems, denoted by SPCA-P1, SPCA-P2 or SPCA-P3 when using the intersection of an l1 ball and an l2 ball, an l1 sphere and an l2 sphere, or an l1 ball and an l2 sphere as constrained set, respectively. We prove the equivalence of the solutions to SPCA-P1 and SPCA-P3, and the solutions to SPCA-P2 and SPCA-P3 are the same in most case. Then by employing the projection method onto the intersection of an l1 ball/sphere and an l2 ball/sphere, we design a gradient projection method (GPSPCA for short) and an approximate Newton algorithm (ANSPCA for short) for SPCA-P1, SPCA-P2 and SPCA-P3 problems, and prove the global convergence of the proposed GPSPCA and ANSPCA algorithms. Finally, we conduct several numerical experiments in MATLAB environment to evaluate the performance of our proposed GPSPCA and ANSPCA algorithms. Simulation results confirm the assertions that the solutions to SPCA-P1 and SPCA-P3 are the same, and the solutions to SPCA-P2 and SPCA-P3 are the same in most case, and show that ANSPCA is faster than GPSPCA for large-scale data. Furthermore, GPSPCA and ANSPCA perform well as a whole comparing with the typical SPCA methods: the l0-constrained GPBB algorithm, the l1-constrained BCD-SPCAl1 algorithm, the l1-penalized ConGradU and Gpowerl1 algorithms, and can be used for large-scale computation.