scikit learn clf.fit / score model accuracy

Question

I'm build a model clf say

clf = MultinomialNB()
clf.fit(x_train, y_train)

then I want to see my model accuracy using score

clf.score(x_train, y_train)

the result was 0.92

My goal is to test against the test so I use

clf.score(x_test, y_test)

This one I got 0.77 , so I thought it would give me the result same as this code below

clf.fit(X_train, y_train).score(X_test, y_test)

This I got 0.54. Can someone help me understand why would 0.77 > 0.54 ?

Answer 1

You must get the same result if x_train, y_train, x_test and y_test are the same in both cases. Here is an example using iris dataset, as you can see both methods get the same result.

>>> from sklearn.naive_bayes import MultinomialNB
>>> from sklearn.cross_validation import train_test_split
>>> from sklearn.datasets import load_iris
>>> from copy import copy
# prepare dataset
>>> iris = load_iris()
>>> X = iris.data[:, :2]
>>> y = iris.target
>>> X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
# model
>>> clf1 = MultinomialNB()
>>> clf2 = MultinomialNB()
>>> print id(clf1), id(clf2) # two different instances
 4337289232 4337289296
>>> clf1.fit(X_train, y_train)
>>> print clf1.score(X_test, y_test)
 0.633333333333
>>> print clf2.fit(X_train, y_train).score(X_test, y_test)
 0.633333333333