Scatter Separability Criterion (SSC) for clustering algorithms

[hristog]

I don't think I've seen references to the Scatter Separability Criterion (SSC) metric anywhere yet - neither in the Issues section here, nor on the mailing list.

Has it been ever considered and if yes, what are the reasons it hasn't made its way into the sklearn module yet?
Would a potential PR be encouraged?

References:

• Dy, J. G., and Brodley, C. E. (2004), "Feature selection for unsupervised learning," Journal of Machine
  Learning Research, 5, 845-889. [Google Scholar link]
• Separability Criteria for Feature Selection [link]

P.S.: Apologies if I've missed relevant discussions and/or implementations which have already addressed the same question.

[jnothman]

I don't think it's been suggested before.

[jnothman]

Sounds interesting.

[tereshin-pmi51]

it's interesting feature.
sorry, I am starting my career in open-source community and i'd like to be a contributor of scikit project=) @jnothman, could you give me chance to write that functionality, please?
Unfortunately I don't know how to reserve that feature correctly and writing here. Sorry if it doesn't work in that way.

[hristog]

@tereshin-pmi51, @jnothman: Sorry, I thought it had been already implied in my original message, that I would like to contribute this feature, if it's considered a potentially valuable addition.
I've used this metric in a couple of MATLAB projects of mine and that's why I was surprised that it wasn't available in sklearn (whilst, I view it as somehow traditional metric in the aforementioned context of evaluating clustering algorithms).

[tereshin-pmi51]

oh, sorry

[glemaitre]

@tereshin-pmi51 There plenty of other scikit-learn issues waiting for you :)

[tereshin-pmi51]

@glemaitre - could you show me link with question what people currently need in scikit? because I can't find currently hot issues...

[jnothman]

GitHub recently establish the "help wanted" label as a standard way to find issues where a project wants help. But I suggest you also find something more straightforward than this. Please read our contributor guide.

@hristog, an unsupervised feature selection approach would be appreciated from my perspective. PR welcome.

[hristog]

@jnothman, thanks for your confirmation. I'll try to get out a PR asap.

[Ryan4Reed]

Hi @hristog, has any progress been made in implementing the scatter separability criterion u mentioned above in 2017?