SMOTE With Non-Continuous Data

[jtsmith2]

I have a dataset with a feature that takes on only integer values (a performance rating, for example). Since SMOTE generates data along a continuous vector between two points, if they had differing values of this feature, the resulting point has a non-integer value. Using a decision tree, it quickly learns that these non-integer values are a good predictor of being in the minority class. For example, I end up with tree splits like: if performance <4 and then the next split is if performance > 3 to a leaf with all minority class data (all generated by SMOTE).

For example, a leaf of all SMOTE data when Performance Rating is 3 < x < 4.

In the original SMOTE paper, the authors suggested a SMOTE-NC for non-continuous data that would use the median of k-nearest neighbors for those non-continuous features. Is there is will to implement this feature? I suppose the user would need to pass the an index list of the non-continuous features over what is currently being passed to the algorithm.

[dvro]

@jtsmith2 idk if we're ought to include SMOTE-NC algorithm considering that we're also not supporting SMOTE for categorical data. So, at least for now we should keep it simple. For your specific problem I'd perform the following workaround:

import numpy as np

# regular smote processing here
# resulting in X, y ...

nc_feat_idxs = [0,1,5]
X[:,nc_feat_idxs] = np.round(X[:,nc_feat_idxs])

Let me know how it works!

[]'s

[glemaitre]

Move this is in new method #105