Does pyGAM support multi-class classification? #196
Comments
|
@zhangxz1123 no. currently pygam does not support multiclass classification. however you can quickly extend pygam to do so. if you have M classes, then you can train M one-vs-all models. when you predict, keep the label for the model that outputs the highest probability. this is a little less efficient, but about the same as the softmax activation. |
|
I was thinking about this as well! Would this work? model__multiclass = OneVsRestClassifier(LogisticGAM())
model__multiclass
# OneVsRestClassifier(estimator=LogisticGAM(callbacks=['deviance', 'diffs', 'accuracy'],
# fit_intercept=True, max_iter=100, terms='auto', tol=0.0001,
# verbose=False),
# n_jobs=1) |
|
@jolespin right, does that work? i havent tried it out. |
|
@dswah I couldn't get it to work: from pygam import LogisticGAM
from sklearn.ensemble import OneVsRestClassifier
base_estimator = LogisticGAM(n_splines=20)
ensemble = OneVsRestClassifier(base_estimator, n_jobs=1)
ensemble.fit(X_iris, y_iris)
model_selection.cross_val_score(ensemble, X=X_iris, y=y_iris, cv=10)
# ---------------------------------------------------------------------------
# AttributeError Traceback (most recent call last)
# ~/anaconda/envs/python3/lib/python3.6/site-packages/sklearn/multiclass.py in _predict_binary(estimator, X)
# 94 try:
# ---> 95 score = np.ravel(estimator.decision_function(X))
# 96 except (AttributeError, NotImplementedError):
# ~/anaconda/envs/python3/lib/python3.6/site-packages/pygam/terms.py in __getattr__(self, name)
# 978
# --> 979 return self._super_get(name)
# 980
# ~/anaconda/envs/python3/lib/python3.6/site-packages/pygam/terms.py in _super_get(self, name)
# 899 def _super_get(self, name):
# --> 900 return super(MetaTermMixin, self).__getattribute__(name)
# 901
# AttributeError: 'LogisticGAM' object has no attribute 'decision_function' |
|
blah, well thats annoying. @jolespin thank you for trying! |
|
this PR (#213) appears to do the trick |
|
Should this be a default for multiclassification case? |
|
@jolespin do you mean that pygam should import scikit-learn? i think pygam should NOT import sklearn because it is such a big library with its various dependencies.... but then perhaps it doesn't make sense to add the |
|
Since this is still open, is it a good approach to have your target split into 'n' 1vsAll class manually and fit 'n' GAMs? I understand it can be tedious for a large number of classes but for example, if you have 3 classes = [1,2,3], you can have 3 GAMs where each GAM is a binary classifier of class 'i' vs. rest=0. This is assuming that splitting your target is simple enough. You could do this manually instead of using OVR from sklearn. |
No description provided.
The text was updated successfully, but these errors were encountered: