L1 feature selection example

l1.py

import numpy as np

from sklearn import linear_model
from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import CountVectorizer


def print_features(coef, names):
    """ Print sorted list of non-zero features/weights. """
    print "\n".join('%s/%.2f' % (names[j], coef[j]) for j in np.argsort(coef)[::-1] if coef[j] != 0)


if (__name__ == '__main__'):
    rand = np.random.mtrand.RandomState(8675309)
    cats = ['rec.sport.baseball', 'sci.crypt']
    data = fetch_20newsgroups(subset='train',
                              categories=cats,
                              shuffle=True,
                              random_state=rand)
    # Smaller C means fewer features selected.
    clf = linear_model.LogisticRegression(penalty='l1', C=.1)
    vec = CountVectorizer()
    X = vec.fit_transform(data.data)
    clf.fit(X, data.target)
    print_features(clf.coef_[0], vec.get_feature_names())

Output is:

clipper/1.43
key/0.84
encryption/0.81
code/0.58
your/0.32
nsa/0.29
uk/0.27
chip/0.25
com/0.19
of/0.18
pgp/0.16
government/0.15
it/0.15
to/0.11
org/0.11
gtoal/0.08
writes/0.08
can/0.05
clinton/0.02
by/0.02
for/0.02
would/0.02
us/0.02
that/0.01
is/0.01
sports/-0.00
in/-0.01
philadelphia/-0.02
go/-0.02
the/-0.03
think/-0.04
fan/-0.06
braves/-0.08
have/-0.08
cubs/-0.09
than/-0.09
games/-0.22
phillies/-0.22
out/-0.24
lines/-0.26
edu/-0.31
subject/-0.42
players/-0.42
team/-0.51
game/-0.53
year/-0.54
he/-0.58
baseball/-1.40