Hi all. Im trying to make a dimensionality reduction of a corpus to visualise in 2D. In doing so, I’m performing an MFCC analysis with 13 dimensions. However, the flattening is arbitrary, meaning that the flattening is completed, giving us a pre-defined order of flattening of those dimensions. This means that the 2D plot would be different if we chose the cluster’s last two dimensions when the final fattening is performed ( this line of thought is given by a linear algebra way of seeing it).

Is there a way to define the order of dimensions or which dimensions use to flatten the corpus?

Also, I would like to know if there is a chance to show or mark tags in the 2D plot.

What did you use for the dimensionality reduction (PCA, UMAP)?

In either case, the reduction is done in a manner that none of the original dimensions are present any more. Meaning that it’s not an “arbitrary order” of the existing ones, but rather a completely new representation. In the case of PCA the first of those dimensions will be the dimension with the most variance, and the second will be the 2nd next amount of variance etc… With UMAP I’m not entirely sure how the dimensional order ir done, but I do believe it is random(ly seeded).

There is a way to determine which dimensions (of your original ones) carry the most variance (a good thread on SVM here) but that may not be ideal for using a 2d plot.

The tags (labels?) can remain attached to each data point regardless of how many dimensions there are. As in, if your first data point (“0” or whatever) has 13d of MFCC data, or 2d of PCA data, it still is point “0”, and if you have a corresponding label (from `fluid.labelset~`

) for “0”, then you can show that just the same.

Well, in this case, I did it with UMAP.

What I would like to achieve, by selecting which dimensions are the final dimensions of the flattening process, is to create various representations of the cluster of data points and see which one accommodates more to a personal view of the characteristics of the sounds.

For example:

A couple of sounds have a distinct timbre, and another cluster of sounds has a tone with high contrast compared to the previously mentioned ones. One would assume there should be some distance or even be located opposite in the 2d mad, yet they are suspiciously close. My idea was to find the 2D flattening presentations that best accommodated my vision.

As for the Tags, I wanted to say that I would like to determine the centroid of a cluster of data points and add a tag to that centroid to evaluate those coordinates with other centroids. I hope it makes sense what I say

What may work is to try different descriptor recipes. It may be worthwhile including pitch as well, and/or trying different combinations of descriptors etc…

Here’s a thread that may be of interest too:

It goes into a bunch of detail about other things, but you can see how similar data has very different representations depending on the descriptors you feed in.

Makes a bit of sense. I guess its worth remembering that the means of the reduction are relative to that newly created space, and not necessarily the original space. You can calculate the means of each “cluster” on the original data, and then run that through the same UMAP projection, and then plot *that* so that way the means are plotting in the same projection. It gets a bit brain puzzley thing about stuff like this.