leeprevost / ordinalclassifier Goto Github PK

@leeprevost Thank you for creating and sharing this work with the community. I have reviewed the code and was unable to locate where I could input the binary class weights, which were calculated using the compute_sample_weight function in scikit-learn. I would appreciate any guidance you can provide on this matter.

Initial version was "hard coded" around 3 class labeling of 0,1,2. Rework to generalize for class labeling.

File "C:\Users\lee\Dropbox\python sandbox\get_yahoo_data\models\ordinal.py", line 249, in fit
raise ValueError("This classifier expects target y to be multiclass. Got type: {}".format(self.y_type_))
ValueError: This classifier expects target y to be multiclass. Got type: binary

Hi, how can I install this package ?

pip install OrdinalClassifier doesn't work.

Thanks in advance for your help!

I think I have a problem when I hand this classifier a y with text based classes (ie. low, medium, high). The default sorter sorts this lexographically (low, high, medium) which makes medium act like a pos class.

sklearn has some PRs for this:
#scikit-learn/scikit-learn#4450
#scikit-learn/scikit-learn#13631

Looks like I need to:

• encode any ordinal informaiton from y
• don't default to "sorted"
• label binarize y and keep separate classes_ array

I already allow for custom ordering.

From testing on diabetes dataset, ordering does have a big effect on precision and recall of pos_class (ie. "high" disease progression in diabetes) so important to get this right.

LP

Classifier does not pickle. Stops with error about dictionary keys.

Allow input to _init to allow for custom ordering of classes possibly by ordring from most important class (postive class) to least important (trivial class). For example, a three class problem where you are trying to classify stocks:

Buy
Hold
Sell

One may want high precision on "buy" class with less focus on recall. But, conversely, high recall on sell class. And less concerned with inaccurate predictions on hold class. So ordering may be:

Buy (pos class)
Sell (neg class)
Hold (neutral class)