1
$\begingroup$

I have a huge dataset with about 1.000.000 matrix entries of size about 300.000 and ran a PCA on them, but the components and the eigenvalues are really small. I am unsure what this means.

I ran an Incremental PCA (from sklearn 0.16.1) on the data with batch size 100. Already for the first increment the values were quite small and then kept decresing with every iteration.

So for the normalized eigenvalues after looking at all the data I get something like this:

array([ 0.01146837, 0.00927029, 0.0067319 , 0.00601581, 0.00507061, 0.00442723, 0.00411958, 0.00397337, 0.00384924, 0.00379049, 0.00339444, 0.00307697, 0.00298565, 0.00279872, 0.00273881, 0.00269498, 0.00264522, 0.00251268, 0.00237168, 0.0021606 , 0.00209025, 0.00203315, 0.00198865, 0.00195354, 0.00192039, 0.001854 , 0.0017387 , 0.00172961, 0.0016596 , 0.00156462, 0.00155386, 0.00150622, 0.00150157, 0.00139544, 0.00134342, 0.00131588])

Can anyone give some intuition what these values mean and why they are so small?

Thank you

$\endgroup$
  • 3
    $\begingroup$ It looks like "the death of 10,000 cuts" (or 1,000,000 cuts) - each vector in your data is the sum of many vectors. Relative difference (subtract min, then divide by overall range, for each value) shows which are more or less significant contributors. Did you uniformly randomly select your "batch"? Has your iterative algorithm converged, perhaps you are taking the cookies out of the oven too early? $\endgroup$ – EngrStudent Jul 22 '15 at 14:40

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.