[FEA] PaCMAP manifold dimension reduction

[ck37]

Is your feature request related to a problem? Please describe.

I would like to have a GPU-accelerated version of the PaCMAP algorithm, a competitor to UMAP. Personally, I would use this with the BERTopic topic-modeling framework but PaCMAP could in theory be used anywhere that algorithms like tSNE or UMAP are used.

Describe the solution you'd like

Similar to cuML.manifold.umap, it would be a python implementation of PaCMAP that runs on GPU.

Additional context

• The PaCMAP GitHub repo is here: https://github.com/YingfanWang/PaCMAP
• The paper is here (JMLR 2021): https://jmlr.org/papers/v22/20-1061.html

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[mikful]

I would also like to see this. I find PaCMAP to generate more useable embeddings, faster and with less parameter optimization.

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[raulcarlomagno]

up 👍

[mvarnold]

Also interested