Support Vector Machines in H2O?

[sarem...@googlemail.com]

Hi guys,

I was wondering if there is any implementation of Support Vector Machines in H2O?
Been looking up Google for some time now, but I couldn't find anything comparable.

If there is no implementation, could you recommend any alternative (I'm using R by the way) that is especially capable of multithreaded SVM?

Thanks for your help,
Sarem

[Erin LeDell]

Hi Sarem,

The SVM algorithm has some inherent scalability challenges and although it's possible we could implement this algo in the future, we have chosen not to implement it thus far.

This post suggests that caret's SVM is multicore (can't verify this). https://geekoverdose.wordpress.com/2014/07/25/svm-classification-example-with-performance-measures-using-r-caret/  This may just mean that it's tuning itself via cross-validation, which is done in parallel, and the actual training of the SVM model is single-threaded...

Best,
Erin

-- 
Erin LeDell Ph.D.
Statistician & Machine Learning Scientist | H2O.ai

[pranas....@algoteq.com]

There are researchers who are dealing with scalable deviations of SVM http://elkingarcia.github.io/Papers/MLDM07.pdf not sure if you're aware about ...

[pra...@orangecap.net]

On Tuesday, December 6, 2016 at 6:23:27 PM UTC+11, pranas....@algoteq.com wrote:
> There are researchers who are dealing with scalable deviations of SVM http://elkingarcia.github.io/Papers/MLDM07.pdf not sure if you're aware about ...

OR

L1-Regularized Distributed Optimization: A Communication-Efficient Primal-Dual Framework

https://arxiv.org/pdf/1502.03508v2.pdf

[dym...@gmail.com]

Hey Sarem,

As Erin mentioned we decided not to implement our own SVM, instead in Sparkling Water we expose Spark's SVM implementation (http://spark.apache.org/docs/latest/mllib-linear-methods.html#linear-support-vector-machines-svms).

This feature is only available through Sparkling Water (and I'm still working on a blog post about it), but if that's ok with you then just download one of the newest versions and you can try it out through FlowUI. I understand this might be a bit inconvenient for you as you're using R, maybe rsparkling could help here?

Regards,
Mateusz

[Erin LeDell]

Mateusz,

That's an interesting idea about using rsparkling to access the SVM. I
haven't tried that yet, but I'd be interested to hear if that would work.

Sarem,

Here is more info on rsparkling: http://spark.rstudio.com/h2o.html

It doesn't look like sparklyr supports SVM via MLlib yet
http://spark.rstudio.com/mllib.html, so rsparkling might be a good solution.

-Erin

[Sarem Seitz]

Wow, thanks for all your efforts to help me. I will check out everything ;)