In [1]:

    

import numpy as np
import pandas as pd
import scipy
import nltk
import sklearn
import random
import re
from sklearn.feature_extraction.text import CountVectorizer,TfidfTransformer
from sklearn.preprocessing import OneHotEncoder
from sklearn.multiclass import OneVsRestClassifier
from sklearn.metrics import f1_score, precision_score, recall_score
from sklearn.linear_model import LogisticRegression, LinearRegression
from sklearn.naive_bayes import GaussianNB
from sklearn.decomposition import PCA, RandomizedPCA
from sklearn.ensemble import RandomForestClassifier


    

In [2]:

    

nltk.download('reuters')
nltk.download('punkt') # needed for tokenization


    

    




[nltk_data] Downloading package reuters to /home/felipe/nltk_data...
[nltk_data]   Package reuters is already up-to-date!
[nltk_data] Downloading package punkt to /home/felipe/nltk_data...
[nltk_data]   Package punkt is already up-to-date!






    
Out[2]:





True

In [3]:

    

dataset = nltk.corpus.reuters


    

In [4]:

    

fileids = dataset.fileids()


    

In [5]:

    

# http://scikit-learn.org/stable/modules/feature_extraction.html#text-feature-extraction
corpus_train = []
corpus_test = []
for fileid in dataset.fileids():
    document = dataset.raw(fileid)
    if re.match('training/',fileid):
        corpus_train.append(document)
    else:
        corpus_test.append(document)


    

In [6]:

    

def preprocessor(string):
    repl = re.sub('<','',string)
    return repl.lower()


    

In [7]:

    

vectorizer = CountVectorizer(
                min_df=10, # tweaking this parameter reduces the length of the feature vector
                strip_accents='ascii',
                preprocessor=preprocessor,
                stop_words='english')


    

In [8]:

    

# need to use both corpuses for fitting because otherwise there may be words that only occur in the
# training set or in the test set
full_corpus = corpus_train + corpus_test
vectorizer.fit(full_corpus)

X_train_counts = vectorizer.transform(corpus_train)
X_test_counts = vectorizer.transform(corpus_test)
X_full_counts = vectorizer.transform(full_corpus)


    

In [9]:

    

transformer = TfidfTransformer()
# again, we need to fit the transformer to all documents (train and test)
transformer.fit(X_full_counts)

X_train_tfidf = transformer.transform(X_train_counts)
X_test_tfidf = transformer.transform(X_test_counts)
X_full_tfidf = transformer.transform(X_full_counts)


    

In [10]:

    

%%time

Y_train = []
Y_test = []

for (idx,fileid) in enumerate(dataset.fileids()):    
    categories = '*'.join(dataset.categories(fileid))

    if re.match('training/',fileid):
        Y_train.append(categories)
    else:
        Y_test.append(categories)

series_train = pd.Series(Y_train)
Y_train_df = series_train.str.get_dummies(sep='*')

series_test = pd.Series(Y_test)
Y_test_df = series_test.str.get_dummies(sep='*')

Y_train = Y_train_df.values
Y_test = Y_test_df.values


    

    




CPU times: user 8.13 s, sys: 77.4 ms, total: 8.21 s
Wall time: 8.1 s

In [23]:

    

%%time

clf = RandomForestClassifier(n_estimators=25)
meta_clf = OneVsRestClassifier(clf,n_jobs=-1)

meta_clf.fit(X_train_tfidf.toarray(),Y_train)


    

    




CPU times: user 1min 10s, sys: 542 ms, total: 1min 10s
Wall time: 1min 56s

In [24]:

    

Y_pred = meta_clf.predict(X_test_tfidf.toarray())


    

In [25]:

    

f1_score(Y_test,Y_pred,average='micro')


    

    
Out[25]:





0.76837593268772808

In [ ]: