Leland McInnes

It exists, and you can use it right now; it isn't necessarily well or prominently placed in any documentation at this time. Depending on how big your distributions are there...

Actually the 1D case just got added to pynndescent (lmcinnes/pynndescent#155), and so should also be available in UMAP with ``metric="wasserstein-1d"`` if you get the current version of pynndescent from github...

I'm working on this currently. Hopefully I'll have the fix merged in soon.

The UMAP class has an attribute ``embedding_`` that contains the relevant coordinates of the points. On Thu, Oct 28, 2021 at 6:37 PM Silvia Pagliarini ***@***.***> wrote: > I am...

The principle is to treat the labels as merely a different view of the data which has a different associated metric. For categorical labels we simply use a categorical metric...

Nothing super obvious springs to mind. The binary metrics like dice are always going to be a little trickier, but conceptually it should all work roughly as expected. I'll try...

If you create a vector where labels are given a positive value and unlabelled points are given a value of -1 then that will be interpreted appropriately when passed as...

UMAP does not predict labels, it merely provides a low dimensional representation. If you want to predict labels you'll need to use a classifier on the low dimensional representation.

It is almost certainly the labels that are contributing the clean class separation. Moreover since UMAP works fundamentally with distances and inter-relationships rather than coordinates it is not really possible...

Numba is currently a pretty hard dependency. They do have some newer releases (and prereleases) available on their own conda channel. Perhaps working with a newer numba may help? It...