See also #4468, #4801 which offer an information-theoretic discretisation. Tree-based discretisation is also suggested there.

It's not clear why you want a one-hot matrix initially rather than ordinal features. "Unary" is another possible encoding.

I think the one-hot encoding makes sense for cross-product features. Ordinal features would be useful to remove noise but not for cross-product features.

Yes but we provide tools to transform ordinal into one-hot

On 10 November 2015 at 19:33, Mathieu Blondel [email protected]
wrote:

I think the one-hot encoding makes sense for cross-product features.
Ordinal features would be useful to remove noise but not for cross-product
features.

—
Reply to this email directly or view it on GitHub
#5778 (comment)
.

Good point. I think I could live with a tool that produces ordinal features then :)

Implementation at #4801

#4801 is not an implementation of what I proposed. Maybe we could have a strategy option (uniform, quantile, mdlp) in a Discretizer class.

MDLP is optionally supervised, iirc.

On 10 November 2015 at 23:29, Mathieu Blondel [email protected]
wrote:

#4801 #4801 is not an
implementation of what I proposed. Maybe we could have a strategy option
(uniform, quantile, mdlp) in a Discretizer class.

—
Reply to this email directly or view it on GitHub
#5778 (comment)
.

@jnothman : Yes, MDLP is supervised, based upon class labels.

I'll be willing to work on #4801 if there is more interest in it. I've gotten more familiar with the scikit-learn source code since I started the PR, and I'm willing to polish the code up.

I also opened up #5003 a long while ago.

Thanks @hlin117. It seems the jury is still out on the advantages of MDLP over simpler methods, or methods reusing existing scikit-learn estimators like trees. As @amueller said there, it needs a motivating example to show that it is useful.

That's fine, I understand, @jnothman.

It seems that @mblondel's discretization description is very similar to a function in R called cut. (cut is a function in R's standard library.)

> vector <- c(0.05, 0.22, 0.33, 0.5, 0.65, 0.99, 0.87)
> cut(vector, breaks=c(0, 0.33, 0.66, 1))
[1] (0,0.33]    (0,0.33]    (0,0.33]    (0.33,0.66] (0.33,0.66] (0.66,1]
[7] (0.66,1]
Levels: (0,0.33] (0.33,0.66] (0.66,1]

The difference is that the user has to designate where the breaks are placed. If the user passes in an integer k for breaks, then it'll bin the points into k intervals, each with the same number of items. (With the chance that one bin will have at most one than another, if the number of items n is not divisible by k.)

Is a PR welcome for this? We can discuss how we could like the class to be structured in the PR.

@hlin117 That would be nice, thanks!

Thanks for the support, @mblondel. I'll work on this PR.

I would start simple (only uniform binning) and add more strategies in other PRs.

Please check the PR in #5825. Thanks!

A discretizer has been merged to a branch (https://github.com/scikit-learn/scikit-learn/tree/discrete). Should be merged to master once some remaining features and an example are added.

Fixed in #9342