quick'n'dirty subsampling

[probcomp]

quick'n'dirty subsampling

[riastradh-probcomp]

vkmvkmvkm: In the future, will we want each model to be analyzed on a (potentially) different randomly chosen subset of rows?

[vkmvkmvkmvkm]

Yes. (If it helps, I'm happy to discuss an interface to models and
meta-models that could make this easy and also help decouple BQL & its
primitives from crosscat.)

On Fri, May 29, 2015 at 6:59 PM, riastradh-probcomp <
[email protected]> wrote:

vkmvkmvkm: In the future, will we want each model to be analyzed on a
different subset of rows?

—
Reply to this email directly or view it on GitHub
#6 (comment)
.

[riastradh-probcomp]

Fixed in 0780f7c.

Selection of rows is not random: it is the first ones in the table. Randomizing it is a separate issue, requires deciding whether to do it nondeterministically or with a fixed seed, and is not what old bayesdb did anyway.

Selection of rows is not per-model: every model shares a common subsampling of rows. Crosscat is not currently set up to allow two models to be trained on two distinct sets of rows.