Supervising training then unlabelled test data problem

[carluqcor]

Hey there,

referring to this section of the documentation, a supervising training using fit and transforming the unllabeled test data could be done. However when I try this approach, I get the following results:

Training using y in fit

Test using transform

This is the UMAP that I'm using. My dataset has (9997, 250, 1) as shape.

reducer = umap.UMAP(
        n_jobs=-1,
        n_components=3,
        n_neighbors=30,
        random_state=0,
    ).fit(data, y=classes.values)

As you can see in the test UMAP, there is a connection path in within every cluster which I can't see happening in the MNIST example. But is this usual?

If further information is needed I will be glad to add it!

Thanks