WrappedClassificationModel() usage

[Luca Zavarella]

I'd like to have a clarification about the functioning of WrappedClassificationModel().

I have a prefitted binary classification model model1 and a function compute_prob that calibrates the probabilities associated to class 1. (model1 has a predict_proba method which isn't calibrated).

Through:
mymodel = WrappedClassificationModel(model1, compute_prob)
a new model is correctly generated (both methods predict and predict_proba work correctly).

It seems though that the predict method uses the same exact method of my original model, while predict_proba uses the wrapped function compute_prob.

1. What I'd like to achieve is obtaining classification labels based on the probabilities computed through compute_prob (1 if probability > treshold, 0 otherwise) directly from the predict method. Is this possible?

2. When using the new wrapped model with SHAP TabularExplainer to generate an explanation (which worked when using my original model), I get an error:

explainer = TabularExplainer(model1,
                                             X_train_df, 
                                             #features=features, 
                                             classes=["Not Fraud","Fraud"])
	
global_explanation_test = explainer.explain_global(X_test_df) 

The error I get is the following:

TypeError                                 Traceback (most recent call last)
<ipython-input-328-0abd05dcccf3> in <module>
----> 1 global_explanation_test = explainer.explain_global(X_test_calibration)
[...]
/anaconda/envs/azureml_py36/lib/python3.6/site-packages/interpret_community/explanation/explanation.py in __init__(self, init_data, eval_data, eval_y_predicted, eval_y_predicted_proba, **kwargs)
	 1234         :type eval_y_predicted_proba: numpy.array or str
	 1235         """
-> 1236         super(_DatasetsMixin, self).__init__(**kwargs)
	 1237         self._init_data = init_data
	 1238         self._eval_data = eval_data

TypeError: object.__init__() takes no parameters

I also tried different eval_functions to make sure the problem doesn't lie specifically in my compute_prob function.

Is there a way to solve the issue and use the wrapped model to generate an explanation?

[Ilya Matiach]

hi @lucazav , you can wrap your model directly, in a class similar to the WrappedClassificationModel(), and define the predict_proba and predict functions in the way you prefer. The model just needs to have these two functions defined to work with the blackbox explainers. The wrap_model function tries to intelligently figure this out (and may use WrappedClassificationModel or others internally), but it can't handle every possible scenario - just the most common ones from the most common frameworks used, so that the input and output follow the scikit-learn standard.

[Ilya Matiach]

@lucazav I looked at the error above but I'm not sure why it is happening. Would you be able to send me a sample notebook with a dummy dataset that reproduces the error, and I can take a look at it and try to fix the issue?

[Ilya Matiach]

@lucazav I also replied on the other issue #457 with an example wrapper if it helps

[Luca Zavarella]

Hi @imatiach-msft,

I used the wrapper example you suggested to me and now I can render the explainability dashboard without errors. Thanks.