Ok, we all laughed about brobdingnagian rat genitals, but for now also a poor statistical choice perfectly suffices to generate a nightmare š±.
In the flagship paper of the
All of Us Research Program published yesterday in Nature, they used UMAP to represent human genetic diversity and populations! That must be among the worst ideas ever in scientific publishing šµāš« - unless you want racists to have a field day.
If not even the scientists understand the statistical methods they are using, how on earth can the general public in the Telegram groups, where those plots will inevitably end up being shared, be trusted to not be mislead by them?
⢠Distances in 2D projections of t-SNE/UMAP plots are not directly interpretable!
⢠Their loss functions are invariant with respect to rotations!
Consequently this visual representation widely exaggerates the distinctiveness, a take-away that every racist will adore! Please share
Understanding UMAP widely, so you and your colleagues know better when using UMAP in your manuscripts!