在scikit-learn中使用SVC分类器进行错误的预测？

Question

I generated my own corpus, so I split into a training text file like this:

POS|This film was awesome, highly recommended
NEG|I did not like this film
NEU|I went to the movies
POS|this film is very interesting, i liked a lot
NEG|the film was very boring i did not like it
NEU|the cinema is big
NEU|the cinema was dark

And for testing I have another text review, which is unlabeled:

I did not like this film

Then I do the following:

import pandas as pd
from sklearn.feature_extraction.text import HashingVectorizer

trainingdata = pd.read_csv('/Users/user/Desktop/training.txt',
                 header=None, sep='|', names=['labels', 'movies_reviews'])


vect = HashingVectorizer(analyzer='word', ngram_range=(2,2), lowercase=True, n_features=7)
X = vect.fit_transform(trainingdata['movies_reviews'])
y = trainingdata['labels']
TestText= pd.read_csv('/Users/user/Desktop/testing.txt',
                     header=None, names=['test_opinions'])
test = vect.transform(TestText['test_opinions'])
from sklearn.svm import SVC
svm = SVC()
svm.fit(X, y)

prediction = svm.predict(test)
print prediction

And the prediction is:

['NEU']

Then something that comes to my mind is why this prediction is wrong?. Does this is a code problem or a feature or a classification algorithm problem?, I tried to play with this and when I remove the last review from the training text file I realize that always is predicting the last element of that file. Any idea of how to fix this problem?.

Answer 1

SVMs are notoriously sensitive to parameter settings. You will need to do a grid search to find the right values. I tried training two kinds of Naive Bayes on your dataset and I got perfect accuracy on the training set:

from sklearn.naive_bayes import *
from sklearn.feature_extraction.text import *

# first option- Gaussian NB
vect = HashingVectorizer(analyzer='word', ngram_range=(2,2), lowercase=True)
X = vect.fit_transform(trainingdata['movies_reviews'])
y = trainingdata['labels']
nb = GaussianNB().fit(X.A,y) # input needs to be dense
nb.predict(X.A) == y

# second option- MultinomialNB (input needs to be positive, use CountingVect instead)
vect = CountVectorizer(analyzer='word', ngram_range=(2,2), lowercase=True)
X = vect.fit_transform(trainingdata['movies_reviews'])
y = trainingdata['labels']
nb = MultinomialNB().fit(X,y)
nb.predict(X.A) == y

In both cases the output is

Out[33]: 
0    True
1    True
2    True
3    True
4    True
5    True
6    True
Name: labels, dtype: bool