luispedro
commented
13 years ago Thank you!
I have merged your request. I am going to use it as a basis for some changes (which I think will improve it): more documentation, adding a test, &c
I will also see whether having the multi class methods propagate the weights is easy or not.
Thanks Luis
On Saturday, October 15, 2011 02:34:29 PM you wrote:
Hello Luis!
milk/supervised/weighted_voting_adaboost.py is the code I was supposed to send. I see it as a generalization of AdaBoost for working with multiple class labels.
A learner for this classifier initialization must be already constructed, e.g.: learner = weighted_voting_ada_learner(100, one_against_one(tree_learner(criterion=neg_z1_loss)))
Maybe I am wrong, but I've discovered that multiclass strategies can't work with weights. For an example, one_against_one.train (milk/supervised/multi.py:114). I think this train() method works with no respect to input weights. And that probably causes my adaboost incorrect learning.
Best regards, Igor
You can merge this Pull Request by running:
git pull https://github.com/ishalyminov/milk master
Or you can view, comment on it, or merge it online at:
https://github.com/luispedro/milk/pull/2
-- Commit Summary --
- Weighted voting adaboost is a generalization of AdaBoost
-- File Changes --
A milk/supervised/weighted_voting_adaboost.py (76)
-- Patch Links --
https://github.com/luispedro/milk/pull/2.patch https://github.com/luispedro/milk/pull/2.diff
Hello Luis!
milk/supervised/weighted_voting_adaboost.py is the code I was supposed to send. I see it as a generalization of AdaBoost for working with multiple class labels.
A learner for this classifier initialization must be already constructed, e.g.: learner = weighted_voting_ada_learner(100, one_against_one(tree_learner(criterion=neg_z1_loss)))
Maybe I am wrong, but I've discovered that multiclass strategies can't work with weights. For an example, one_against_one.train (milk/supervised/multi.py:114). I think this train() method works with no respect to input weights. And that probably causes my adaboost incorrect learning.
Best regards, Igor