tf-idf feature weights using sklearn.feature_extraction.text.TfidfVectorizer

Question

this page: http://scikit-learn.org/stable/modules/feature_extraction.html mentions:

As tf–idf is a very often used for text features, there is also another class called TfidfVectorizer that combines all the option of CountVectorizer and TfidfTransformer in a single model.

then I followed the code and use fit_transform() on my corpus. How to get the weight of each feature computed by fit_transform()?

I tried:

In [39]: vectorizer.idf_
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-39-5475eefe04c0> in <module>()
----> 1 vectorizer.idf_

AttributeError: 'TfidfVectorizer' object has no attribute 'idf_'

but this attribute is missing.

Thanks

Answer 1

Since version 0.15, the tf-idf score of each feature can be retrieved via the attribute idf_ of the TfidfVectorizer object:

from sklearn.feature_extraction.text import TfidfVectorizer
corpus = ["This is very strange",
          "This is very nice"]
vectorizer = TfidfVectorizer(min_df=1)
X = vectorizer.fit_transform(corpus)
idf = vectorizer.idf_
print dict(zip(vectorizer.get_feature_names(), idf))

Output:

{u'is': 1.0,
 u'nice': 1.4054651081081644,
 u'strange': 1.4054651081081644,
 u'this': 1.0,
 u'very': 1.0}

---

As discussed in the comments, prior to version 0.15, a workaround is to access the attribute idf_ via the supposedly hidden _tfidf (an instance of TfidfTransformer) of the vectorizer:

idf = vectorizer._tfidf.idf_
print dict(zip(vectorizer.get_feature_names(), idf))

which should give the same output as above.

Answer 2

See also this on how to get the TF-IDF values of all the documents:

feature_names = tf.get_feature_names()
doc = 0
feature_index = X[doc,:].nonzero()[1]
tfidf_scores = zip(feature_index, [X[doc, x] for x in feature_index])
for w, s in [(feature_names[i], s) for (i, s) in tfidf_scores]:
    print w, s

this 0.448320873199
is 0.448320873199
very 0.448320873199
strange 0.630099344518

#and for doc=1
this 0.448320873199
is 0.448320873199
very 0.448320873199
nice 0.630099344518

I think the results are normalized by document:

>>>0.4483208731992+0.4483208731992+0.4483208731992+0.6300993445182 0.9999999999997548