Is training a semi or supervised dimensionality reduced space with UMAP using multi-label targets supported & known to yield meaningful results (with respect to the unsupervised embedding)?
The documentation shows we can use multi-class labels as in the fashion MNIST dataset where each clothing object belongs to one clothing type (enumerated as an integer or -1 if unlabeled). I was wondering if the same technique works on embedding news article objects given a one-hot-encoded label of their known categories?
