In [1]:

    

'''load dataset'''
from sklearn.datasets import fetch_20newsgroups
train = fetch_20newsgroups(subset='train')
test = fetch_20newsgroups(subset='test')


    

In [2]:

    

'''feature extraction using different ngram analyzers'''
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.svm import LinearSVC
ngrams = [(1, 1, 2), (1, 1, 3), (1, 1, 4), (1, 2, 2), (1, 2, 3)]
train_Xs, test_Xs = [], []
from pandas import DataFrame, Series, Index
frame = DataFrame(columns=['vectorizer', 'ngram_min', 'ngram_max', 'num_feature', 'min_df',
                           'train_time', 'model', 'score'], )
string = train.data[0][:50]
from fabric.colors import red, blue
print red('Example text: %s' % string)
index = 0
for ngram_min in range(1, 4):
    for ngram_max in range(ngram_min, 4):
        for min_df in range(1, 3):
            vect = CountVectorizer(ngram_range=(ngram_min, ngram_max), min_df=min_df)
            anal = vect.build_analyzer()
            analyze_eg = str(anal(string))
            print blue('ngram min=%d max=%d min_df=%d' % (ngram_min, ngram_max, min_df))
            print '\tresult:%s\n' % (analyze_eg)
            model = LinearSVC()
            frame = frame.append(DataFrame({'vectorizer': vect,
                                            'ngram_min': ngram_min,
                                            'ngram_max': ngram_max,
                                            'min_df': min_df,
                                            'model': model,
                                            },
                                            index=Index([index])))
            index += 1

frame


    

    




Example text: From: lerxst@wam.umd.edu (where's my thing)
Subjec
ngram min=1 max=1 min_df=1
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec']

ngram min=1 max=1 min_df=2
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec']

ngram min=1 max=2 min_df=1
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec', u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec']

ngram min=1 max=2 min_df=2
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec', u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec']

ngram min=1 max=3 min_df=1
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec', u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec', u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']

ngram min=1 max=3 min_df=2
	result:[u'from', u'lerxst', u'wam', u'umd', u'edu', u'where', u'my', u'thing', u'subjec', u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec', u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']

ngram min=2 max=2 min_df=1
	result:[u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec']

ngram min=2 max=2 min_df=2
	result:[u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec']

ngram min=2 max=3 min_df=1
	result:[u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec', u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']

ngram min=2 max=3 min_df=2
	result:[u'from lerxst', u'lerxst wam', u'wam umd', u'umd edu', u'edu where', u'where my', u'my thing', u'thing subjec', u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']

ngram min=3 max=3 min_df=1
	result:[u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']

ngram min=3 max=3 min_df=2
	result:[u'from lerxst wam', u'lerxst wam umd', u'wam umd edu', u'umd edu where', u'edu where my', u'where my thing', u'my thing subjec']







    
Out[2]:






  

    

      

      
min_df
      
model
      
ngram_max
      
ngram_min
      
num_feature
      
score
      
train_time
      
vectorizer
    
  
  

    

      
0 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 1
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
1 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 1
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
2 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
3 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
4 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
5 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 1
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
6 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 2
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
7 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 2
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
8 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 2
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
9 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 2
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
10
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 3
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
11
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 3
      
 NaN
      
 NaN
      
 NaN
      
 CountVectorizer(analyzer=word, binary=False, c...
    
  

In [3]:

    

'''Run experiments'''
import time
num_train, num_test = 2000, 500
for index, exp in frame.iterrows():
    start_time = time.time()
    train_X, test_X = train.data[:num_train], test.data[:num_test]
    train_Y, test_Y = train.target[:num_train], test.target[:num_test]
    vect = exp['vectorizer']
    vect.fit(train_X)
    vect.fit(test_X)
    train_X, test_X = vect.transform(train_X), vect.transform(test_X)
    model = exp['model']
    model.fit(train_X, train_Y)
    frame.ix[index, 'num_feature'] = train_X.shape[1]
    frame.ix[index, 'train_time'] = time.time() - start_time
    frame.ix[index, 'score'] = model.score(test_X, test.target[:test_X.shape[0]])

frame


    

    
Out[3]:






  

    

      

      
min_df
      
model
      
ngram_max
      
ngram_min
      
num_feature
      
score
      
train_time
      
vectorizer
    
  
  

    

      
0 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 1
      
 1
      
  16316
      
 0.614
      
 5.340934
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
1 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 1
      
 1
      
   6461
      
 0.598
      
 4.803072
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
2 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 1
      
  90265
      
 0.636
      
 12.13506
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
3 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 1
      
  17796
      
 0.626
      
 7.965464
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
4 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 1
      
 193684
      
 0.642
      
  17.5893
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
5 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 1
      
  24442
      
 0.632
      
 12.89869
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
6 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 2
      
  73949
      
 0.552
      
 6.732035
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
7 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 2
      
 2
      
  11335
      
 0.498
      
 5.102484
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
8 
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 2
      
 177368
      
 0.548
      
  14.7479
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
9 
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 2
      
  17981
      
 0.484
      
 9.869578
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
10
      
 1
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 3
      
 103419
      
 0.472
      
 8.157876
      
 CountVectorizer(analyzer=word, binary=False, c...
    
    

      
11
      
 2
      
 LinearSVC(C=1.0, class_weight=None, dual=True,...
      
 3
      
 3
      
   6646
      
  0.39
      
  5.16128
      
 CountVectorizer(analyzer=word, binary=False, c...
    
  

In [4]:

    

'''Plot result'''
import matplotlib.pyplot as plt
fig, axs = plt.subplots(1, 3)
fig.set_figwidth(15)
ax = axs[0]
ax.scatter(frame['num_feature'].values, frame['score'])
ax.set_xlim(0)
ax.set_xlabel('num_feature')
ax.set_ylabel('score (accuracy)')
ax = axs[1]
ax.scatter(frame['min_df'].values, frame['score'])
ax.set_xlabel('min_df')
ax = axs[2]
ax.scatter(frame['ngram_max'].values, frame['score'])
ax.set_xlabel('ngram_max')


    

    
Out[4]:





<matplotlib.text.Text at 0x26c01550>