[BUG] lightgbm.basic.booster classifier was recognized as regressor

[omrialo]

Hi,
My model, which is object of type lightgbm.basic.Booster, was recognized by deepchecks as a regressor, while it is a classifier.
When I used the method model.dump_model(), the objective was:
'objective': 'binary sigmoid:1'

Could you please look into that?

[noamzbr]

Hi @omrialo. The checks and suites in deepchecks expect the model to provide the methods of the sklearn api for models, as detailed in our Supported Models guide. The way deepchecks recognizes classification models is by checking if they have a predict_proba method, which deepchecks uses to compute the predicted class probabilities for all classes. Models with only a predict method are assumed to be regression models.

In the case of lightgbm.basic.Booster, it has only a predict method, that either produces the class probability of the second class for binary classification, or of all classes for multiclass. The solution for that is to write a short model wrapper object that will make sure the model provides to deepchecks the api it expects. For example, in your case the wrapper can be:

import numpy as np

class ModelWrapper:
    
    def __init__(self, model):
        self.model = model
    
    def predict(self, data):
        return (self.model.predict(data) > 0.5) * 1
    
    def predict_proba(self, data):
        predicted_class_1 = self.model.predict(data).reshape((-1, 1))
        return np.concatenate([1-predicted_class_1, predicted_class_1], axis=1)

Alternatively, you can train your model using the lightgbm sklearn api. For example, the lightgbm.LGBMClassifier object already implements the sklearn api and will work out of the box with deepchecks.