I was reading the CART book recently and actually, another way of using
class weights, is to use them explicitly as priors when computing the
criterion values. That is, instead of using p(j | t) = N_j(t) / N(t) as the
probability estimate of class j within node t, define p(j | t) = p(j, t) /
p(t), where p(j, t) = pi(j) N_j(t) / N_j and p(t) = \sum_j p(j, t), where
pi(j) is the prior for class j. (And then proceed as usual for computing
the criterion.)

On 28 January 2014 14:36, Arnaud Joly notifications@github.com wrote:

Can't this be achieved by applying compute_class_weight() for each output
variable separately? Is there any literature or any other resource where I
can refer to?

I don't have any article at hand about it. But basically, you would have
to generalize the single output behavior to multiple-output. This would
need to be done at the Criterion level. From a user perspective,
class_weight for a single output could be obtained by modifying the
sample_weight as described by @amueller https://github.com/amueller.

Reply to this email directly or view it on GitHubhttps://github.com//issues/2129#issuecomment-33478287
.

Uh oh!

Implement class_weights parameter for Random Forests #2129

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Implement `class_weights` parameter for Random Forests #2129