ExponentialNormalization doesn't seem to work if data hasn't been pre-standardized somehow.
Closed this issue · 1 comments
jdurbin commented
Why? It is rank based, so it shouldn't care what the range of input values is. First showed up in egrfusion dataset that was unnormalized from MedBook. Even weirder, it only shows up when doing classification with wmClassify. During cross-validation it seems to work fine, with or without pre-normalization. Why is that?
jdurbin commented
Problem related to calling filter with a dataset that had a fake class attribute added and no ID.
Solution was to perform filtering BEFORE this operation.
Might want to revisit code at some point to try to understand in greater detail exactly the mechanism of failure and see if a defensive protection against this can be added.