PCA matrix data compression help

Hi,

I'm making a neural network for classification(newff or patternnet) and I have a input matrix 400x500 (rows x column) and a target vector 1x500 with [zeros ones] my true/false.

Which PCA algorithm and how I should use on my input matrix to get a matrix 100x500 or 10x500 or 5x500 (data compression) but also to use my target matrix with zeros&ones on this data?

Thank you :)

Answer 1

[Greg Heath]

For classification, choose the dimensions in the direction of greatest class separation.

This is not guaranteed using PCA which chooses the dimensions with the largest variances.

For a detailed explanation, search comp.ai.neural-nets and/or comp.soft-sys.matlab with

heath cigar

heath parallel cigar

PLS (Partial-Least-Squares) is more appropriate.

Hope this helps.

Greg