0
$\begingroup$

I have a dataset in which each sample has 2D matrix data, e.g., 50x1000 size matrix. Each of these 50 rows has different meanings. I convert this to 1*50,000 size data and use svm for classification. But now I want to find out the most significant rows for classification. I checked Sequential Feature Selection of matlab, but the problem is, it discards and takes some columns from row 1, some columns from row 2 etc. But I want to use each row as a group for feature selection so that either I take that whole row or not. Does anyone know any algorithms for this purpose?

$\endgroup$

1 Answer 1

1
$\begingroup$

Use PCA to collapse the 50 dimension down to e.g. 3 or 4 dimensions (or the number of PCs for which their eigenvalues > 1). The information in the 50 will remain within the smaller dimension set, and in this fashion you also won't be using different features from the set of 50 for each sample. PCA dimension reduction will employ a fair use of all the information form the original set. Also, I would stack all the samples on top of one another (thousands of rows with 50 columns) and run PCA on the correlation matrix of the 50 columns. Don't split the samples and run PCA on each (50x1000 array) sample separately.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.