Closed jianshu93 closed 7 months ago
Its not faster in our experiments. See https://github.com/ThrunGroup/BanditPAM/issues/175
Because we do not have distance functions, but currently operate on precompupted distance matrixes only, this is currently out-of-scope for this package, but this may, of course, change in the future. Because the available implementation of BanditPAM performed worse for us (despite being already C++), it is not of a high priority.
We would appreciate the contribution of such algorithms as long as they do not add major external dependencies (to avoid dependency hell). I would suggest to first add:
These are quite similar but use uniform sampling, although my impression that FasterPAM will still outperform these if you want high accuracy. I do not think the BadingPAM authors ever compared to these baseline method...
Note there is a newer version of BanditPAM to be published at NeurIPS 2023. The preprint is here: https://arxiv.org/pdf/2310.18844.pdf
But I am not keen of that paper either, because it
Hence I am a bit reluctant to promote their work by implementing it.
Dear rust-k-mediods team,
This is amazing implementation and many thanks for this. I noticed an even faster one, which was based on fastPAM but 4 times faster. It is not an exact solution but approximate solution. However, it is as accuracy as the exact solution in practice. Check here: https://proceedings.neurips.cc/paper/2020/file/73b817090081cef1bca77232f4532c5d-Paper.pdf
Any thoughts and potential to also implement this? It would be even useful for very large dataset.
Thanks,
Jianshu