large-Scale Supervised Sparse Principal Component Analysis

From statwiki
Revision as of 00:54, 14 August 2013 by Z47xu (talk | contribs) (Block Coordinate Ascent Algorithm)
Jump to: navigation, search

Introduction

The sparse PCA is a variant of the classical PCA, which assumes sparsity in the feature space. It has several advantages such as easy to interpret, and works for really high-dimensional data. The main issue about sparse PCA is that it is computationally expensive. Many algorithms have been proposed to solve the sparse PCA problem, and the authors introduced a fast block coordinate ascent algorithm with much better computational complexity.

1 Drawbacks of Existing techniques

Existing techniques include ad-hoc methods(e.g. factor rotation techniques, simple thresholding), greedy algorithms, SCoTLASS, the regularized SVD method, SPCA, the generalized power method. These methods are based on non-convex optimization and they don't guarantee global optimum.

A semi-definite relaxation method called DSPCA can guarantee global convergence and has better performance than above algorithms, however, it is computationally expensive.

2 Contribution of this paper

This paper solves DSPCA in a computationally easier way, and hence it is a good solution for large scale data sets. This paper applies a block coordinate ascent algorithm with computational complexity [math]O(\hat{n^3})[/math], where [math]\hat{n}[/math] is the intrinsic dimension of the data. Since [math]\hat{n}[/math] could be very small compared to the dimension [math]n[/math] of the data, this algorithm is computationally easy.

Primal problem

The sparse PCA problem can be formulated as [math]max_x \ x^T \Sigma x - \lambda \| x \|_0 : \| x \|_2=1[/math].

This is equivalent to [math]max_z \ Tr(\Sigma Z) - \lambda \sqrt{\| Z \|_0} : Z \succeq 0, Tr Z=1, Rank(Z)=1[/math].

Replacing the [math]\sqrt{\| Z \|_0}[/math] with [math]\| Z \|_1[/math] and dropping the rank constraint gives a relaxation of the original non-convex problem:

[math]\phi = max_z Tr (\Sigma Z) - \lambda \| Z \|_1 : Z \succeq 0[/math], [math]Tr(Z)=1 \qquad (1)[/math] .

Fortunately, this relaxation approximates the original non-convex problem to a convex problem.

Here is an important theorem used by this paper:

Theorem(2.1) Let [math]\Sigma=A^T A[/math] where [math]A=(a_1,a_2,......,a_n) \in {\mathbb R}^{m \times n}[/math], we have [math]\psi = max_{\| \xi \|_2=1}[/math] [math]\sum_{i=1}^{n} (({a_i}^T \xi)^2 - \lambda)_+[/math]. An optimal non-zero pattern corresponds to the indices [math]i[/math] with [math]\lambda \lt (({a_i}^T \xi)^2-\lambda)_+[/math]

Block Coordinate Ascent Algorithm

There is a row-by-row algorithm applied to the problems of the form [math]min_X \ f(X)-\beta \ log(det X): \ L \leq X \leq U, X \succ 0[/math].

Problem (1) can be written as [math]{\frac 1 2} {\phi}^2 = max_X \ Tr \Sigma X - \lambda \| X \|_1 - \frac 1 2 (Tr X)^2: X \succeq 0 \qquad (2)[/math] .

In order to apply the row by row algorithm, we need to add one more term [math]\beta \ log(det X)[/math] to (2) where [math]\beta\gt 0[/math] is a penalty parameter.

That is to say, we address the problem [math]\ max_X \ Tr \Sigma X - \lambda \| X \|_1 - \frac 1 2 (Tr X)^2 + \beta \ log(det X): X \succeq 0 \qquad (3)[/math]

By matrix partitioning, we could obtain the sub-problem:

[math]\phi = max_{x,y} \ 2(y^T s- \lambda \| y \|_1) +(\sigma - \lambda)x - {\frac 1 2}(t+x)^2 + \beta \ log(x-y^T Y^{\dagger} y ):y \in R(Y) \qquad (4)[/math].

By taking the dual of (4), the sub-problem can be simplified to be

[math] {\phi}^' = min_{u,z} {\frac 1 {\beta z}} u^T Yu - \beta (log z) + {\frac 1 2} (c+ \beta z)^2 : z\gt 0, \| u-s \|_\infty \leq \lambda [/math]

Since [math] \beta [/math] is very small, and we want to avoid large value of [math] z [/math], we could change variable [math]r=\beta z[/math], then the optimization problem become

[math] {\phi}^' - \beta (log \beta) = min_{u,r} {\frac 1 r} u^T Yu - \beta (log r) + {\frac 1 2} (c+r)^2 : r\gt 0, \| u-s \|_\infty \leq \lambda [/math]


Here is the algorithm:


algorithm.jpg


Convergence and complexity

1. The algorithm is guaranteed to converge to the global optimizer.

2. The complexity for the algorithm is [math]O(K \hat{n^3})[/math], where [math]K[/math] is the number of sweeps through columns (fixed, typically [math]K=5[/math]), and [math](\hat{n^3})[/math] is the intrinsic dimension of the data points.