difference between coef_ and feature_log_prob_ in multinomial naive bayes?

Question

The below code represnets sklearn multinomial naive bayes.

import numpy as np
from sklearn.naive_bayes import MultinomialNB
X = np.random.randint(5, size=(10, 100))
y=np.random.randint(2,size=(10,))
clf = MultinomialNB()
clf.fit(X, y)

Then I want to find out the important features in my model and in sklearn documentation we have two parameters namely.

 feature_log_prob_ : array, shape (n_classes, n_features)
 Empirical log probability of features given a class, P(x_i|y).

coef_ : array, shape (n_classes, n_features)
Mirrors feature_log_prob_ for interpreting MultinomialNB as a linear model.

Then If I try to print both attributes

print(clf.feature_log_prob_.shape)  // giving (2,100)
print(clf.coef_.shape)         // giving (1,100)

But when my classes are more than two then both attributes giving the same results.

what is the difference between two above attributes?

Answer 1

In standard binary classification coef_ gives you the probability of observing the "success" category. In multinomial case, coef_ returns probabilities of observing each of the outcomes, i.e for all classes it will return prob score.