Mapping Competitive Biotech Landscapes with Embeddings(rxdatalab.com)

2 pointsby nnmgقبل 10 أشهر1 comments

Mapping Competitive Biotech Landscapes with Embeddings

https://rxdatalab.com/research/biotech-competitive-landscape-embeddings/

1 comments

PaulHouleقبل 10 أشهر

I know UMAP has gotten to be the standard these days but I don't really like it. In particular you see these cusp structures that strike me as a "bad smell" in high dimension data visualizations -- I imagine in these cases some points lie on a manifold around the edge of a fold which is perpendicular to the plane so something that really should be a blob looks like a line and worse yet two sides of the blob get projected on top of each other; this was something I struggled with with my PhD drawing Poincare sections until I realized that what was going on.

I worked on a visualization product which used t-SNE which I thought was the first cluster visualization that really worked. Our version was tuned up so that you got nice blobs instead of those nasty lines. Sometimes I see cuspy t-SNEs but I think if your parameters tuned right you get better and more insightful visualizations than UMAP.