Better handling of NaNs

[nicodv]

np.unique is used for encoding data values to integers. However, numpy currently treats every np.NaN as a unique value, creating many categories. (See: numpy/numpy#2111)

Solution is to check with np.isnan so that we can just ignore all NaNs when encoding. The encoder then simply assigns all NaNs to the -1 (i.e. "I don't know this value") category.

[allefeld]

@nicodv, could you expand on this a bit, or give a code example? How does one "check with np.isnan" when using np.unique? Also, a remark in the documentation of np.unique would be useful.

[nicodv]

@allefeld , have a look at the commit that fixes this issue: 6fd7c98

It basically filters out NaNs in the encoding and makes sure they get a -1 value.