0
$\begingroup$

I'm trying to gain an intuition for the 2nd dimension in the spectral embedding of an S-shaped dataset as in this example: enter image description here

The 1st dimension seems to neatly capture the local similarity between points, but is there a similar physical interpretation for the 2nd dimension? More specifically, why do both red AND blue points show low values for this dimension?

I am wondering whether these dimensions can be interpreted in a similar way as done for the embedding of hand-written digits:

enter image description here

$\endgroup$
  • $\begingroup$ Consider the sample in the bottom left of the LLE. It's not thick at all. It didn't capture thickness on the y axis. That was just a pretty story to make it sound cool. $\endgroup$ – Anony-Mousse May 30 at 5:20
0
$\begingroup$

The first dimension yields nthr axis along which to split to get two partitions.

The second dimension is for splitting into 3 or 4 clusters. Try running k-means on this projection with k=3 or k=4 and you should get some decent results.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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