UMAP experiments #1
Description
-
cool!! this will help everyone easily understand how to call UMAP
-
general suggestion: if you use an example from e.g. the UMAP authors, give an attribution. For example, the real data sets are well identified. Otherwise, it will be assumed that the whole experiment design is yours. E.g. I understand that the 4 dimension RGB data experiment design is Qirui's own. it's very cool!
-
if we have an interesting example for 1 algorithm, it will be good to replicate it for the other algorithms, just to illustrate the differences (keep this in mind)
-
regarding UMAP: we want to focus on examples where there is a low dimensional structure that UMAP (or another algorithm can uncover). This is not the case for the colors example, because there the intrinsic dimension d=4, and all the embeddings were in dimension m < 4. So UMAPcould show some features, but could never recover the data.
Hence, let's run a UMAP experiment with d=2 instead of 4 and with one or more of the data sets created by Yujia and see what happens (for example, a rectangle and a circle). When this is done, let me know, I will propose a new experiment that is specific to UMAP, it highlights the properties of min_dist.