UMAP

Topological data analysis

• https://github.com/lmcinnes/umap

https://www.youtube.com/watch?v=YPJQydzTLwQ

Related to t SNE? https://openreview.net/forum?id=B8a1FcY0vi

Tutorials and usage tips

• Nikolay Oskolkov: How Exactly UMAP Works
  • How to Program UMAP from Scratch
• Pay attention to the initialization: Kobak2021initialization.

Libraries

• UMAP.js

Drawbacks and criticisms

• https://www.nature.com/articles/s41586-023-06957-x/figures/2 from AllOfUs2024genomic
  • The UMAP algorithm, by design, exaggerates the distinctiveness of the most frequent ancestries, a message that can be misinterpreted by the public.
• Biologists, stop putting UMAP plots in your papers

Lior Pachter has been a vocal critic. For instance, see this thread and Chari2021specious.

Tips

By Filipi:

If you see such a weird umap, it may be creating disconnected components in the knn network, which will be expelled from the center of the viz. Recommendations are: (1) increase the number of neighbors if you have a lot of data. (2) sample your data! check if a sample of ~100k points results in something meaningful. (3) If you see a blob without much structure, you may need to increase epochs. (4) If you are using a older version of cuml, use random initialization as spectral initialization was broken.If nothing works, it could be another cuml bug. (5) Try umap-learn with a smaller sample to double check.