Binary matrix as input

[scikit-mine]

Binary matrix as input

[remiadon]

UPDATE

Considering only basic transactional datasets, MultiLabelBinarizer gives access to a .classes_ attributes. The problem is that it is an attribute of the object, it does not transit within data --> sklearn MultiLabelBinarizer produces np.array or scipy.sparse matrics, but they are not labels

In the context of Pattern mining we absolutely need the labels : theses are our beloved symbols$

>>> from skmine.preprocessing import MulitLabelBinarizer
>>> mb = MultiLabelBinarizer(sparse_output=True)
>>> D = pd.Series([  # SLIM takes a pd.Series as input  
>>>    ['bananas', 'milk'],  
>>>     ['milk', 'bananas', 'cookies'],  
>>>   ['cookies', 'butter', 'tea'],  
>>>    ['tea'],   
>>>    ['milk', 'bananas', 'tea'],  
>>>  ])
>>> tab_D = mb.fit_transform(D)  # scipy sparse matrix
>>> tab_D[:2].todense()
matrix([[1, 0, 0, 1, 0],
        [1, 0, 1, 1, 0]])
>>> mb.classes_
array(['bananas', 'butter', 'cookies', 'milk', 'tea'], dtype=object)

Note that pandas can build a DataFrame from sparse matrices, as introduced in version 0.25

>>> tab_D = pd.DataFrame.sparse.from_spmatrix(tab_D, columns=mb.classes_)
>>> tab_D
   bananas  butter  cookies  milk  tea
0        1       0        0     1    0
1        1       0        1     1    0
2        0       1        1     0    1
3        0       0        0     0    1
4        1       0        0     1    1

>>> tab_D.dtypes
bananas    Sparse[int64, 0]
butter     Sparse[int64, 0]
cookies    Sparse[int64, 0]
milk       Sparse[int64, 0]
tea        Sparse[int64, 0]
dtype: object

My solution:

• reimplement skearn MultiLabelBinarizer to make its transform function return a pandas.DataFrame with some explicit columns (our symbols)
• propose this in sklearn as a feature request. If accepted, we will use it as implemented in sklearn to make sure we don't reimplement things twice

[remiadon]

see
e4b467b