应用scikit-learn做文本分类

文本挖掘的paper没找到统一的benchmark，只好自己跑程序，走过路过的前辈如果知道20newsgroups或者其它好用的公共数据集的分类（最好要所有类分类结果，全部或取部分特征无所谓）麻烦留言告知下现在的benchmark，万谢！

嗯，说正文。20newsgroups官网上给出了3个数据集，这里我们用最原始的20news-19997.tar.gz。

分为以下几个过程：

• 加载数据集
• 提feature
• 分类
  • Naive Bayes
  • KNN
  • SVM

• 聚类

说明：  scipy官网 上有参考，但是看着有点乱，而且有bug。本文中我们分块来看。

Environment： Python 2.7 + Scipy （scikit-learn）

1.加载数据集

从20news-19997.tar.gz下载数据集，解压到scikit_learn_data文件夹下，加载数据，详见code注释。

 #first extract the 20 news_group dataset to /scikit_learn_data
 from sklearn.datasets import fetch_20newsgroups
 #all categories
 #newsgroup_train = fetch_20newsgroups(subset='train')
 #part categories
 categories = ['comp.graphics',
  'comp.os.ms-windows.misc',
  'comp.sys.ibm.pc.hardware',
  'comp.sys.mac.hardware',
  'comp.windows.x'];
 newsgroup_train = fetch_20newsgroups(subset = 'train',categories = categories);

可以检验是否load好了：

 #print category names
 from pprint import pprint
 pprint(list(newsgroup_train.target_names))

结果：

['comp.graphics',
 'comp.os.ms-windows.misc',
 'comp.sys.ibm.pc.hardware',
 'comp.sys.mac.hardware',
 'comp.windows.x']

2. 提feature：

刚才load进来的newsgroup_train就是一篇篇document，我们要从中提取feature，即词频啊神马的，用fit_transform

Method 1. HashingVectorizer，规定feature个数

 #newsgroup_train.data is the original documents, but we need to extract the
 #feature vectors inorder to model the text data
 from sklearn.feature_extraction.text import HashingVectorizer
 vectorizer = HashingVectorizer(stop_words = 'english',non_negative = True,
                                n_features = 10000)
 fea_train = vectorizer.fit_transform(newsgroup_train.data)
 fea_test = vectorizer.fit_transform(newsgroups_test.data);


 #return feature vector 'fea_train' [n_samples,n_features]
 print 'Size of fea_train:' + repr(fea_train.shape)
 print 'Size of fea_train:' + repr(fea_test.shape)
 #11314 documents, 130107 vectors for all categories
 print 'The average feature sparsity is {0:.3f}%'.format(
 fea_train.nnz/float(fea_train.shape[0]*fea_train.shape[1])*100);

结果：

Size of fea_train:(2936, 10000)
Size of fea_train:(1955, 10000)
The average feature sparsity is 1.002%

因为我们只取了10000个词，即10000维feature，稀疏度还不算低。而实际上用TfidfVectorizer统计可得到上万维的feature，我统计的全部样本是13w多维，就是一个相当稀疏的矩阵了。

**************************************************************************************************************************

上面代码注释说TF-IDF在train和test上提取的feature维度不同，那么怎么让它们相同呢？有两种方法：

Method 2. CountVectorizer+TfidfTransformer

让两个CountVectorizer共享vocabulary：

 #----------------------------------------------------
 #method 1:CountVectorizer+TfidfTransformer
 print '*************************\nCountVectorizer+TfidfTransformer\n*************************'
 from sklearn.feature_extraction.text import CountVectorizer,TfidfTransformer
 count_v1= CountVectorizer(stop_words = 'english', max_df = 0.5);
 counts_train = count_v1.fit_transform(newsgroup_train.data);
 print "the shape of train is "+repr(counts_train.shape)

 count_v2 = CountVectorizer(vocabulary=count_v1.vocabulary_);
 counts_test = count_v2.fit_transform(newsgroups_test.data);
 print "the shape of test is "+repr(counts_test.shape)

 tfidftransformer = TfidfTransformer();

 tfidf_train = tfidftransformer.fit(counts_train).transform(counts_train);
 tfidf_test = tfidftransformer.fit(counts_test).transform(counts_test);

结果：

*************************

CountVectorizer+TfidfTransformer
*************************
the shape of train is (2936, 66433)
the shape of test is (1955, 66433)

Method 3. TfidfVectorizer

让两个TfidfVectorizer共享vocabulary：

 #method 2:TfidfVectorizer
 print '*************************\nTfidfVectorizer\n*************************'
 from sklearn.feature_extraction.text import TfidfVectorizer
 tv = TfidfVectorizer(sublinear_tf = True,
                                     max_df = 0.5,
                                     stop_words = 'english');
 tfidf_train_2 = tv.fit_transform(newsgroup_train.data);
 tv2 = TfidfVectorizer(vocabulary = tv.vocabulary_);
 tfidf_test_2 = tv2.fit_transform(newsgroups_test.data);
 print "the shape of train is "+repr(tfidf_train_2.shape)
 print "the shape of test is "+repr(tfidf_test_2.shape)
 analyze = tv.build_analyzer()
 tv.get_feature_names()#statistical features/terms

结果：

*************************
TfidfVectorizer
*************************
the shape of train is (2936, 66433)
the shape of test is (1955, 66433)

此外，还有sklearn里封装好的抓feature函数，fetch_20newsgroups_vectorized

Method 4. fetch_20newsgroups_vectorized

但是这种方法不能挑出几个类的feature，只能全部20个类的feature全部弄出来：

 print '*************************\nfetch_20newsgroups_vectorized\n*************************'
 from sklearn.datasets import fetch_20newsgroups_vectorized
 tfidf_train_3 = fetch_20newsgroups_vectorized(subset = 'train');
 tfidf_test_3 = fetch_20newsgroups_vectorized(subset = 'test');
 print "the shape of train is "+repr(tfidf_train_3.data.shape)
 print "the shape of test is "+repr(tfidf_test_3.data.shape)

结果：

*************************
fetch_20newsgroups_vectorized
*************************
the shape of train is (11314, 130107)
the shape of test is (7532, 130107)

3. 分类

3.1 Multinomial Naive Bayes Classifier

见代码&comment，不解释

 ######################################################
 #Multinomial Naive Bayes Classifier
 print '*************************\nNaive Bayes\n*************************'
 from sklearn.naive_bayes import MultinomialNB
 from sklearn import metrics
 newsgroups_test = fetch_20newsgroups(subset = 'test',
                                      categories = categories);
 fea_test = vectorizer.fit_transform(newsgroups_test.data);
 #create the Multinomial Naive Bayesian Classifier
 clf = MultinomialNB(alpha = 0.01)
 clf.fit(fea_train,newsgroup_train.target);
 pred = clf.predict(fea_test);
 calculate_result(newsgroups_test.target,pred);
 #notice here we can see that f1_score is not equal to 2*precision*recall/(precision+recall)
 #because the m_precision and m_recall we get is averaged, however, metrics.f1_score() calculates
 #weithed average, i.e., takes into the number of each class into consideration.

注意我最后的3行注释，为什么f1≠2*（准确率*召回率）/（准确率+召回率）

其中，函数calculate_result计算f1：

 def calculate_result(actual,pred):
     m_precision = metrics.precision_score(actual,pred);
     m_recall = metrics.recall_score(actual,pred);
     print 'predict info:'
     print 'precision:{0:.3f}'.format(m_precision)
     print 'recall:{0:0.3f}'.format(m_recall);
     print 'f1-score:{0:.3f}'.format(metrics.f1_score(actual,pred));

3.2 KNN：

 ######################################################
 #KNN Classifier
 from sklearn.neighbors import KNeighborsClassifier
 print '*************************\nKNN\n*************************'
 knnclf = KNeighborsClassifier()#default with k=5
 knnclf.fit(fea_train,newsgroup_train.target)
 pred = knnclf.predict(fea_test);
 calculate_result(newsgroups_test.target,pred);

3.3 SVM:

 ######################################################
 #SVM Classifier
 from sklearn.svm import SVC
 print '*************************\nSVM\n*************************'
 svclf = SVC(kernel = 'linear')#default with 'rbf'
 svclf.fit(fea_train,newsgroup_train.target)
 pred = svclf.predict(fea_test);
 calculate_result(newsgroups_test.target,pred);

结果：

*************************

Naive Bayes
*************************
predict info:
precision:0.764
recall:0.759
f1-score:0.760
*************************
KNN
*************************
predict info:
precision:0.642
recall:0.635
f1-score:0.636
*************************
SVM
*************************
predict info:
precision:0.777
recall:0.774
f1-score:0.774

4. 聚类

 ######################################################
 #KMeans Cluster
 from sklearn.cluster import KMeans
 print '*************************\nKMeans\n*************************'
 pred = KMeans(n_clusters=5)
 pred.fit(fea_test)
 calculate_result(newsgroups_test.target,pred.labels_);

结果：

*************************
KMeans
*************************
predict info:
precision:0.264
recall:0.226
f1-score:0.213

本文全部代码下载：在此

貌似准确率好低……那我们用全部特征吧……结果如下：

*************************
Naive Bayes
*************************
predict info:
precision:0.771
recall:0.770
f1-score:0.769
*************************
KNN
*************************
predict info:
precision:0.652
recall:0.645
f1-score:0.645
*************************
SVM
*************************
predict info:
precision:0.819
recall:0.816
f1-score:0.816
*************************
KMeans
*************************
predict info:
precision:0.289
recall:0.313
f1-score:0.266