[extension] supervised UMAP for multi-label classification ?

[mglowacki100]

Is there a natural way to extend supervised UMAP to multi-label classification? Of course there is naive approach to transform multi-label into multiclass, but I wonder if there are other approaches possible/makes sense e.g. :

1. transform multi-lable into mulitple binary classifications problems
2. train umap model for every binary classification
3. combine umap models into one

Maybe there are other approaches?

[lmcinnes]

In principle you can build a custom metric (say, similar to jaccard) over the multi-labels, and use that are the target_metric parameters for a surpervised UMAP. You could also try using ParametricUMAP in it's semi-supervised mode and use an appropriate NN architecture for multi-label classification for the classifier portion.