0

I have a feature matrix composed of frequency responses (in dB) from individual acoustic events. Frequencies in the columns, events in the rows and the matrix is the response

The responses decrease with frequency, since the sensor at any given location is more sensitive to low frequency than high frequencies.

I am interested in the loadings since I want to see which frequency components contribute the most to the principle components.

The problem that I have is that when I center and normalize my data, the variances explained change drastically from 90% (expected) to 10%.

Is there any rule of thumb as to which type of normalization to do when such a thing is happening: the data has the same scale, the same units and are all collected using the same sensor (which has a skewed sensitivity to the low frequencies due to attenuation).

If someone can attach any reference as well that would be great! thanks in advance for reading!

Tino D
  • 11
  • 3
  • PCA on centered, standardized variables is PCA based on their *correlation.* – whuber Jul 12 '21 at 13:48
  • @whuber I did not get this point, I know to use the correlation matrix when the variables have different scales/units. But when it comes to covariance ? So can I not center my data ? (I am not using the svd apporach, rather the eigenvalue decomposition) – Tino D Jul 12 '21 at 14:00

0 Answers0