[Bug] Problem with angular metric

[aluo-x]

Given data X, with size [m, n], the fit_transform results differ depending on if the data is pre-normalized (so each row is unit-norm). However the expected result is that the angular metric should remain robust to the norm.

[hyhuang00]

Thanks for reporting. Are you using the default initialization option (with PCA?) In that case, the initialization of the embedding will be based on the pre-normalized data, so it's expected that the result will be different. To disable this setting, you can explicitly set init=random in the fit_transform() method. For reproducibility, you may also want to explicitly set the random state. A snippet is provided below:

import pacmap
reducer = pacmap.PaCMAP(random_state=12345678, distance="angular")
X = your_data()
embedding = reducer.fit_transform(X, init="random")

[aluo-x]

Ah that makes sense, it seems like PCA should be changed to perform normalization as well.

I've "fixed" the bug by always pre-normalizing the data. I've noticed some other issues that seem to be bugs, will write a report at a later point in time once ICLR is over.

[hyhuang00]

Happy to hear your problem is solved and you're using PaCMAP for your research :D Feel free to report any additional problems at any time, and I will try to respond ASAP.