MLARAM bug " return numpy.array(numpy.matrix(allranks))" raises ValueError: matrix must be 2-dimensional

[simon-m]

Hello,

when running the following code, I get the error mentioned in the title:

from skmultilearn.neurofuzzy import MLARAM
mam = MLARAM(vigilance=0.9, threshold=0.02, neurons=[])
from sklearn.datasets import make_multilabel_classification
x, y = make_multilabel_classification(sparse = True, n_labels = 5,
  return_indicator = 'sparse', allow_unlabeled = False)
mam.fit(x.todense(), y.todense())
mam.predict(x.todense())

Full traceback:

Traceback (most recent call last):
  File "C:\Program Files\WinPython-32bit-3.6.1.0Qt5\python-3.6.1\lib\site-packages\IPython\core\interactiveshell.py", line 2881, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-5-1c3b03743990>", line 1, in <module>
    mam.predict(x.todense())
  File "C:\Program Files\WinPython-32bit-3.6.1.0Qt5\python-3.6.1\lib\site-packages\skmultilearn\neurofuzzy\MLARAMfast.py", line 141, in predict
    ranks = self.predict_proba(X)
  File "C:\Program Files\WinPython-32bit-3.6.1.0Qt5\python-3.6.1\lib\site-packages\skmultilearn\neurofuzzy\MLARAMfast.py", line 233, in predict_proba
    return numpy.array(numpy.matrix(allranks))
  File "C:\Program Files\WinPython-32bit-3.6.1.0Qt5\python-3.6.1\lib\site-packages\numpy\matrixlib\defmatrix.py", line 274, in __new__
    raise ValueError("matrix must be 2-dimensional")
ValueError: matrix must be 2-dimensional

Also, note that MLARAM does not support sparse matrices contrary to what is mentioned in the comments. As an aside, in the same method, the test

if len(X) == 0:
    return

fails on sparse matrices and should probably be replaced by something like

        if scipy.sparse.issparse(X):
            if X.getnnz() == 0:
                return
        elif len(X) == 0:
            return

regards,

Simon

[simon-m]

Hello,
I added support for sparse X and y for MLARAM. However, being a complete novice when it comes to working on a shared codebase, I'd rather ask:
Is it appropriate to ask for a pull request now?
Simon

[ChristianSch]

Yes, please do. If there is something off we can review your work. It can always be improved in the remote branch. We then approve your request if everything's good and we tested that it works as intented.
Thanks for your work! Seeing forward to your pull request, cheers.

[niedakh]

i've merged your PRs but it seems that the todense() approach still fails, mostly because neuron vectors become matrices instead of arrays if todense() is used.

@simon-m which representation is the best for a neuron's vector? array, matrix or sparse matrix?

[simon-m]

Well, so far the code seems to rely on operations that cannot be applied to sparse matrices. This PR only adds some todense() where needed to avoid crashes. Thus, until the code is rewritten in a way sparse matrices can be processed, sparse matrices are probably not suited.

From what I see, there is a great deal of numpy.squeeze(numpy.asarray(a.todense())) so this argues for the use of arrays instead of matrices.

Cheers

[niedakh]

@simon-m, ok so I went that way, using arrays internally for now if dense representation is passed. this is fixed in master.

[ghiander]

Hello,

I am trying to use the classifier with a dense matrix (n_samples, n_features) that was converted from a Pandas DataFrame. I tried to use several functions such as "DataFrame.values" and "numpy.matrix(DataFrame)" with no success. The classifier raises the error "ValueError: matrix must be 2-dimensional".

[ghiander]

I forgot to say that I can actually train the classifier using a DataFrame converted using "numpy.matrix(DataFrame)" but it does not work when it comes to predict a test matrix