Does h2o.splitFrame function randomize - or not?

[Ogukku]

I'd like to get a final, complete answer about splitFrame(). In my experience, splitframe seems to "cut" the data into consecutive partitions.

However the following two links suggest that the h2o.splitFrame function introduces randomization as part of the splitting process:

https://groups.google.com/forum/m/#!searchin/h2ostream/Splitframe/h2ostream/zDgBPCb4wkY

http://stackoverflow.com/q/43463178/6088414

Yet, the link below by Arno Candel suggests that splitFrame only "cuts" data into consecutive partitions:

https://stats.stackexchange.com/questions/168480/inverse-progression-for-training-validation-data-during-training-with-h2o

So which one is it?

Thanks!

[Tom Kraljevic]

[ Final, complete answers can change in software! ]

It used to cut.  And now it’s a random row-by-row “coin flip” (comparison of a threshold against a per-row value of runif).

This is the git commit that converted the behavior in Oct. 2015.

https://github.com/h2oai/h2o-3/commit/f2b73c87eb12494998a529360f14da5a2bb11a38

This is the relevant file in the master branch (search for splitFrame):

https://github.com/h2oai/h2o-3/blame/master/h2o-r/h2o-package/R/frame.R

And these are the release branches (as well as the master branch, and therefore future releases) that contain the random behavior.  So it’s been a while.

$ git branch -r --contains f2b73c87eb12494998a529360f14da5a2bb11a38 | grep rel-

  origin/rel-tibshirani

  origin/rel-tukey

  origin/rel-turan

  origin/rel-turchin

  origin/rel-turin

  origin/rel-turing

  origin/rel-turing-8a

  origin/rel-turnbull

  origin/rel-tutte

  origin/rel-tverberg

  origin/rel-ueno

I also added a comment about this on the older post by Arno.

Tom

--
You received this message because you are subscribed to the Google Groups "H2O Open Source Scalable Machine Learning - h2ostream" group.
To unsubscribe from this group and stop receiving emails from it, send an email to h2ostream+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.