OnVsRestClassifier gives 0 accuracy - python-2.7

Question

i am trying to solve a multilabel classification problem as
from sklearn.preprocessing import MultiLabelBinarizer
traindf = pickle.load("traindata.pkl","rb"))
X = traindf['Col1']
X=MultiLabelBinarizer().fit_transform(X)
y = traindf['Col2']
y= MultiLabelBinarizer().fit_transform(y)
Xtrain, Xvalidate, ytrain, yvalidate = train_test_split(X, y, test_size=.5)
from sklearn.linear_model import LogisticRegression
clf = OneVsRestClassifier(LogisticRegression(penalty='l2', C=0.01)).fit(Xtrain,ytrain)
print "One vs rest accuracy: %.3f" % clf.score(Xvalidate,yvalidate)
in this way, i always get 0 accuracy. Please point out if i am doing something wrong. i am new to multilabel classification. Here is what my data looks like
Col1 Col2
asd dfgfg [1,2,3]
poioi oiopiop [4]
EDIT
Thanks for your help #lejlot. I think i am getting the hang of it. Here is what i tried
import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import SGDClassifier
from sklearn.multiclass import OneVsRestClassifier
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
tdf = pd.read_csv("mul.csv", index_col="DocID",error_bad_lines=False)
print tdf
so my input data looks like
DocID Content Tags
1 abc abc abc [1]
2 asd asd asd [2]
3 abc abc asd [1,2]
4 asd asd abc [1,2]
5 asd abc qwe [1,2,3]
6 qwe qwe qwe [3]
7 qwe qwe abc [1,3]
8 qwe qwe asd [2,3]
so this is just some test data i created. then i do
text_clf = Pipeline([
('vect', TfidfVectorizer()),
('clf', SGDClassifier(loss='hinge', penalty='l2',
alpha=1e-3, n_iter=5, random_state=42)),
])
t=TfidfVectorizer()
X=t.fit_transform(tdf["Content"]).toarray()
print X
this gives me
[[ 1. 0. 0. ]
[ 0. 1. 0. ]
[ 0.89442719 0.4472136 0. ]
[ 0.4472136 0.89442719 0. ]
[ 0.55247146 0.55247146 0.62413987]
[ 0. 0. 1. ]
[ 0.40471905 0. 0.91444108]
[ 0. 0.40471905 0.91444108]]
then
y=tdf['Tags']
y=MultiLabelBinarizer().fit_transform(y)
print y
gives me
[[0 1 0 0 1 1]
[0 0 1 0 1 1]
[1 1 1 0 1 1]
[1 1 1 0 1 1]
[1 1 1 1 1 1]
[0 0 0 1 1 1]
[1 1 0 1 1 1]
[1 0 1 1 1 1]]
here i am wondering why there are 6 column? shouldn't there be only 3?
anyway, then i also created a test data file
sdf=pd.read_csv("multest.csv", index_col="DocID",error_bad_lines=False)
print sdf
so this looks like
DocID Content PredTags
34 abc abc qwe [1,3]
35 asd abc asd [1,2]
36 abc abc abc [1]
i have the PredTags column to check for accuracy. So finally i fit and predict as
clf = OneVsRestClassifier(LogisticRegression(penalty='l2', C=0.01)).fit(X,y)
predicted = clf.predict(t.fit_transform(sdf["Content"]).toarray())
print predicted
which gives me
[[1 1 1 1 1 1]
[1 1 1 0 1 1]
[1 1 1 0 1 1]]
Now, how do i know which tags are being predicted? How can i check the accuracy against my PredTags column?
Update
Thanks a lot #lejlot :) i also manged to get the accuracy as follows
sdf=pd.read_csv("multest.csv", index_col="DocID",error_bad_lines=False)
print sdf
predicted = clf.predict(t.fit_transform(sdf["Content"]).toarray())
print predicted
ty=sdf["PredTags"]
ty = [map(int, list(_y.replace(',','').replace('[','').replace(']',''))) for _y in ty]
yt=MultiLabelBinarizer().fit_transform(ty)
Xt=t.fit_transform(sdf["Content"]).toarray()
print Xt
print yt
print "One vs rest accuracy: %.3f" % clf.score(Xt,yt)
i just had to binarize the test set prediction column as well :)

Answer 1

The actual problem is the way you work with text, you should extract some kind of features and use it as text representation. For example you can use bag of words representation, or tfidf, or any more complex approach.
So what is happening now? You call multilabelbinarizer on list of strings thus, scikit-learn creates a set of all iterables in the list... leading to the set of letters representation. So for example
from sklearn.preprocessing import MultiLabelBinarizer
X = ['abc cde', 'cde', 'fff']
print MultiLabelBinarizer().fit_transform(X)
gives you
array([[1, 1, 1, 1, 1, 1, 0],
[0, 0, 0, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 0, 1]])
| | | | | | |
v v v v v v v
a b _ c d e f
Consequently classification is nearly impossible as this does not capture any meaning of your texts.
You could do for example a Count Vectorization (bag of words)
from sklearn.feature_extraction.text import CountVectorizer
print CountVectorizer().fit_transform(X).toarray()
gives you
[[1 1 0]
[0 1 0]
[0 0 1]]
| | |
v | v
abc | fff
v
cde
Update
Finally, to make predictions with labels, and not their binarization you need to store your binarizer thus
labels = MultiLabelBinarizer()
y = labels.fit_transform(y)
and later on
clf = OneVsRestClassifier(LogisticRegression(penalty='l2', C=0.01)).fit(X,y)
predicted = clf.predict(t.fit_transform(sdf["Content"]).toarray())
print labels.inverse_transform(predicted)
Update 2
If you only have three classes then the vector should have 3 elements, yours have 6 so check what you are passing as "y", there is probably some mistake in your data
from sklearn.preprocessing import MultiLabelBinarizer
MultiLabelBinarizer().fit_transform([[1,2], [1], [3], [2]])
gives
array([[1, 1, 0],
[1, 0, 0],
[0, 0, 1],
[0, 1, 0]])
as expected.
My best guess is that your "tags" are also strings thus you actually call
MultiLabelBinarizer().fit_transform(["[1,2]", "[1]", "[3]", "[2]"])
which leads to
array([[1, 1, 1, 0, 1, 1],
[0, 1, 0, 0, 1, 1],
[0, 0, 0, 1, 1, 1],
[0, 0, 1, 0, 1, 1]])
| | | | | |
v v v v v v
, 1 2 3 [ ]
And these are your 6 classes. Three true ones, 2 "trivial" classes "[" and "]" which are present always and also nearly trivial class "," which appears for every object beleonging to more than one class.
You should convert your tags to actual lists first, for example by
y = [map(int, list(_y.replace(',','').replace('[','').replace(']',''))) for _y in y]