Merging to numpy arrays with different column dime

问题:

For a machine learning task I am looking for a way to merge two feature matrices, with different dimensions, so that I can feed them both to an estimator. I cannot use the scipy merging methods since these require compatible shapes. I can use the numpy merging methods, but that goes wrong when I actually try to split the array for cross validation. The error looks like this:

 Traceback (most recent call last):
  File "C:\Users\Ano\workspace\final_submission\src\linearSVM.py", line 50, in <module>
    result = ridge(train_text,train_labels,test_set,train_state,test_state)
  File "C:\Users\Ano\workspace\final_submission\src\Algorithms.py", line 90, in ridge
    x_train, x_test, y_train, y_test = cross_validation.train_test_split(train, labels, test_size = 0.2, random_state = 42)
  File "C:\Python27\lib\site-packages\sklearn\cross_validation.py", line 1394, in train_test_split
    arrays = check_arrays(*arrays, **options)
  File "C:\Python27\lib\site-packages\sklearn\utils\validation.py", line 211, in check_arrays
    % (size, n_samples))
ValueError: Found array with dim 77946. Expected 2

The reason that this error occurs have I found in another stackoverflow question thread:Concatenate sparse matrices in Python using SciPy/Numpy. Apparently np.vstack/hstack create two matrix objects, which caused my error.

The shapes I am dealing with:

(77946, 63677)

(77946, 55)

Basically, I am looking for a way to append those 55 extra features per sample from the second matrix to the features in the first matrix.

I also tried to create a numpy array with the appropriate dimensions and simply fill it with the feature matrices, but even creating that matrix gave me a memory error. I tried to convert it to a sparse matrix, but that didn't work either. Perhaps I am doing something wrong there?

new_matrix = sparse.csr_matrix(np.zeros((77946,63727)))
new_matrix[:,0:63676] = big_feature_matrix
new_matrix[:,63677:63727] = small_feature_matrix

Update So tried out Jaime's solution but it gave me an error:

Code involved

def feature_extraction(train,test,train_small,test_small):


    vectorizer = TfidfVectorizer(min_df = 3,strip_accents = "unicode",ngram_range = (1,2))

    cv = CountVectorizer(strip_accents = "unicode",analyzer = "word",token_pattern = r'\w{1,}')


    print("fitting Vectorizer")
    vectorizer.fit(train)
    train_small = cv.fit_transform(train_state)
    test_small = cv.transform(test_state)
    print("transforming text")
    train = vectorizer.transform(train)
    test = vectorizer.transform(test)

    new_train = sparse.hstack((train, train_small),
                                 format='csr')
    new_test = sparse.hstack((test, test_small),
                                 format='csr')


    return new_train,new_test

Full traceback

Traceback (most recent call last):
  File "C:\Users\Ano\workspace\final_submission\src\linearSVM.py", line 50, in <module>
    result = ridge(train_text,train_labels,test_set,train_small,test_small)
  File "C:\Users\Ano\workspace\final_submission\src\Algorithms.py", line 89, in ridge
    train,test = feature_extraction(train,test,train_small,test_small)
  File "C:\Users\Ano\workspace\final_submission\src\Preprocessing.py", line 109, in feature_extraction
    format='csr')
  File "C:\Python27\lib\site-packages\scipy\sparse\construct.py", line 423, in hstack
    return bmat([blocks], format=format, dtype=dtype)
  File "C:\Python27\lib\site-packages\scipy\sparse\construct.py", line 523, in bmat
    raise ValueError('blocks[%d,:] has incompatible row dimensions' % i)
ValueError: blocks[0,:] has incompatible row dimensions

The train sets have the same dimensions as before. The test sets have less samples (42157).

Update

Jaime's solution, did actually work, I just messed up when I loaded in the files, thank you for all your help!

回答1:

You can use scipy.sparse.hstack:

new_matrix = scipy.sparse.hstack((big_feature_matrix, small_feature_matrix),
                                 format='csr')