deflation Method for Penalized Matrix Decomposition Sparse PCA

From statwiki
Jump to: navigation, search

In the penalized matrix decomposition proposed by Witten, Tibshirani and Hastie<ref name="WTH2009">Daniela M. Witten, Robert Tibshirani, and Trevor Hastie. (2009) "A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis". Biostatistics, 10(3):515–534.</ref>, after the penalized vectors [math]\,\textbf{v}_k[/math] and [math]\,\textbf{u}_k[/math] and the constant [math]\,d_k[/math] have been determined, the data matrix [math]\,\textbf{X}^k[/math] is deflated using the following formula:

[math]\textbf{X}^{k+1} = \textbf{X}^k - d_k\textbf{u}_k\textbf{v}_k^T[/math]

The penalized matrix decomposition can be used to obtain a version of sparse PCA. In this case,

[math]\,\textbf{u}_k = \frac{\textbf{X}^k\textbf{v}_k}{\|\textbf{X}^k\textbf{v}_k\|_2}[/math]

and

[math]\,\textbf{d}_k = \textbf{u}^T_k\textbf{X}^k\textbf{v}_k = \frac{\textbf{v}^T_k\textbf{X}^{kT}\textbf{X}^k\textbf{v}_k}{\|\textbf{X}^k\textbf{v}_k\|_2} = \frac{{\|\textbf{X}^k\textbf{v}_k\|^2_2}}{\|\textbf{X}^k\textbf{v}_k\|_2} = {\|\textbf{X}^k\textbf{v}_k\|_2}.[/math]

Then,

[math]\textbf{X}^{k+1} = \textbf{X}^k - {\|\textbf{X}^k\textbf{v}_k\|_2}\frac{\textbf{X}^k\textbf{v}_k\textbf{v}_k^T}{\|\textbf{X}^k\textbf{v}_k\|_2} = \textbf{X}^k - \textbf{X}^k\textbf{v}_k\textbf{v}_k^T = \textbf{X}^k(I - \textbf{v}_k\textbf{v}_k^T).[/math]

So if [math]\| \textbf{v}_k \|_2 = 1[/math] then the deflation method begin used for the penalized sparse PCA is the projection deflation method.

References

<references/>