FastOPTICS in sklearn.cluster

[DiTo97]

Describe the workflow you want to enable

Is there a roadmap to add the FastOPTICS algorithm, [1], to the sklearn.cluster code base that already supports OPTICS?

[1] 2013, J. Schneider, M. Vlachos, _Fast Parameterless Density-based Clustering via Random Projections

Describe your proposed solution

The solution would be to combine what has already been done for the base OPTICS algorithm, combined with the existing code base for random projections and the Johnson-Lindenstrauss bound in sklearn.random_projection, to implement FastOPTICS.

The implementation in the data mining library ELKI (albeit in Java) could be used as an inspiration.

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[adrinjalali]

On its own the paper doesn't cut our inclusion criteria. But if we see it as a new solver to OPTICS, and the implementation not being too complex, we could think about including it.

Would also need how this plays with the work being done on HDBSCAN: main...hdbscan (cc @Micky774 )

Note that we can NOT use ELKI as inspiration since it's GPL. We went down that road when working on OPTICS, and we needed to implement using the paper, rather than another implementation (which didn't end up being a bad idea anyway).

[DiTo97]

Hi @adrinjalali,

Sorry for the GPL constraint, I am not familiar with ELKI and didn't check the license.

As for the selection criteria, I guess the one that the paper doesn't cut is having 200+ citations, am I right? From a few experiments with ELKI, I would say it really speeds OPTICS up, especially when the data is in very high dimensions, but I also noticed how the paper doesn't provide a quantitative analysis on how close it actually is to the base OPTICS in different data scenarios

[Veghit]

@adrinjalali what's our criteria? (beginner here trying to learn)

[DiTo97]

Hi @Veghit,

I think they're listed here