When using Pandas DataFrames

[mbernico]

_add_shadows_get_imps() fails when X is pandas rather than numpy

Pandas DF can no longer be sliced as
x_cur = np.copy(X[:, x_cur_ind])

x_cur = np.copy(X.as_matrix()[:, x_cur_ind])
OR
x_cur = np.copy(X.ix[:, x_cur_ind])

I'd recommend testing/casting dataframes to numpy arrays in _fit

[danielhomola]

Yepp atm boruta expects a numpy array for X, but this is made explicit in the docstring of fit():
X : array-like, shape = [n_samples, n_features]
The training input samples.

If you feel this is an important issue, please add this to the fit and I'll review your changes.

Oh you did, that's wonderful, cheers!

[mbernico]

The examples show pandas going in. I suppose it would be as easy to just update the user doc to show them to only send numpy. I built a 'pandas check' but that has the unfortunate side effect of adding a dependency. It appears that's how sklearn handles it as well though. Toss up, I'll leave you to decide which you like better :)

[danielhomola]

Hi Mike,

Yepp, I wanted it to have a scikit learn interface, so kinda instinctively stuck with the numpy input as sklearn does.. I added a warning to the examples as you recommended, and renamed boruta_py2 to boruta_py_plus.. Also left in your sanity check for pandas dataframes jsut in case. Pandas is pretty common now, it's not a major dependency issue imo..

Thanks again for your valuable input, really appreciate it!

cheers,
Dan