0
$\begingroup$

I want to do a classification on PC scores. I have a $400$ dimensional matrix, e.g. $2000\times 400$ ($2000$ number of samples and $400$ dimensions). I first apply PCA on it and take it to 3D, i.e. $2000\times 3$. There are $5$ types of classes in this data. When I go to PC space, these classes gets separated but this amount of separability is not enough for my purpose. I'd like to pass data through a non-linear/linear function to increase this superability. I applied LDA but the results was surprisingly worse. What type of function do you recommend to increase the separability? Thanks

$\endgroup$
1
  • $\begingroup$ Could you explain specifically what you mean by "increase separability"? How do you measure "separability"? $\endgroup$
    – whuber
    Feb 25 at 14:50

2 Answers 2

0
$\begingroup$

If you want to increase separability, you shouldn't do dimension reduction. After PCA, you can keep 400 dimensions. Only when you want to visualize your data, use the first three dimensions.

$\endgroup$
1
  • $\begingroup$ Thanks for your response. Ok, ... just assume these 3*2000 are my original data and I want to increase separability. What to do? $\endgroup$ May 19, 2016 at 4:07
0
$\begingroup$

Linear dimensionality reduction, like PCA, does not increase the separability. You should try to separate your data before the PCA.

Non-linear dimensionality reduction may or may not increase the separability, but I don't see any reason to do it after PCA.

$\endgroup$
1
  • $\begingroup$ Thanks for your response. Ok, ... just assume these 3*2000 are my original data and I want to increase separability. What to do? $\endgroup$ May 19, 2016 at 4:07

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.