如何访问稀疏矩阵元素？

Question

type(A)
<class 'scipy.sparse.csc.csc_matrix'>
A.shape
(8529, 60877)
print A[0,:]
  (0, 25)   1.0
  (0, 7422) 1.0
  (0, 26062)    1.0
  (0, 31804)    1.0
  (0, 41602)    1.0
  (0, 43791)    1.0
print A[1,:]
  (0, 7044) 1.0
  (0, 31418)    1.0
  (0, 42341)    1.0
  (0, 47125)    1.0
  (0, 54376)    1.0
print A[:,0]
  #nothing returned

Now what I don't understand is that A[1,:] should select elements from the 2nd row, yet I get elements from the 1st row via print A[1,:] . Also, print A[:,0] should return the first column but I get nothing printed. Why?

Answer 1

A[1,:] is itself a sparse matrix with shape (1, 60877). This is what you are printing, and it has only one row, so all the row coordinates are 0.

For example:

In [41]: a = csc_matrix([[1, 0, 0, 0], [0, 0, 10, 11], [0, 0, 0, 99]])

In [42]: a.todense()
Out[42]: 
matrix([[ 1,  0,  0,  0],
        [ 0,  0, 10, 11],
        [ 0,  0,  0, 99]], dtype=int64)

In [43]: print(a[1, :])
  (0, 2)    10
  (0, 3)    11

In [44]: print(a)
  (0, 0)    1
  (1, 2)    10
  (1, 3)    11
  (2, 3)    99

In [45]: print(a[1, :].toarray())
[[ 0  0 10 11]]

You can select columns, but if there are no nonzero elements in the column, nothing is displayed when it is output with print :

In [46]: a[:, 3].toarray()
Out[46]: 
array([[ 0],
       [11],
       [99]])

In [47]: print(a[:,3])
  (1, 0)    11
  (2, 0)    99

In [48]: a[:, 1].toarray()
Out[48]: 
array([[0],
       [0],
       [0]])

In [49]: print(a[:, 1])


In [50]:

The last print call shows no output because the column a[:, 1] has no nonzero elements.

Answer 2

To answer your title's question using a different technique than your question's details:

csc_matrix gives you the method .nonzero() .

Given:

>>> import numpy as np
>>> from scipy.sparse.csc import csc_matrix
>>> 
>>> row = np.array( [0, 1, 3])
>>> col = np.array( [0, 2, 3])
>>> data = np.array([1, 4, 16])
>>> A = csc_matrix((data, (row, col)), shape=(4, 4))

You can access the indices poniting to non-zero data by:

>>> rows, cols = A.nonzero()
>>> rows
array([0, 1, 3], dtype=int32)
>>> cols
array([0, 2, 3], dtype=int32)

Which you can then use to access your data, without ever needing to make a dense version of your sparse matrix:

>>> [((i, j), A[i,j]) for i, j in zip(*A.nonzero())]
[((0, 0), 1), ((1, 2), 4), ((3, 3), 16)]

Answer 3

If it is for calculating TFIDF score using TfidfTransformer , yu can get the IDF by tfidf.idf_ . Then the sparse array name, say 'a', a.toarray().

toarray returns an ndarray; todense returns a matrix. If you want a matrix, use todense ; otherwise, use toarray .

Answer 4

I fully acknowledge all the other given answers. This is simply a different approach.

To demonstrate this example I am creating a new sparse matrix:

from scipy.sparse.csc import csc_matrix
a = csc_matrix([[1, 0, 0, 0], [0, 0, 10, 11], [0, 0, 0, 99]])
print(a)

Output:

(0, 0)  1
(1, 2)  10
(1, 3)  11
(2, 3)  99

To access this easily, like the way we access a list, I converted it into a list.

temp_list = []
for i in a:
    temp_list.append(list(i.A[0]))

print(temp_list)

Output:

[[1, 0, 0, 0], [0, 0, 10, 11], [0, 0, 0, 99]]

This might look stupid, since I am creating a sparse matrix and converting it back, but there are some functions like TfidfVectorizer and others that return a sparse matrix as output and handling them can be tricky. This is one way to extract data out of a sparse matrix.