Optionally encode the features
I am developing the "linear" (in fact, quadratic) regressor here: !106 (merged)
The only problem is that the trick, position, number of times lead etc are all numeric, and while it is good for the regression tree, this is not optimal for the linear regressor. It needs one-hot encoded factors, and ordered factors must be encoded too.
There are two options: 0. Do not encode;
- Encode "factor <= modality" for each modality;
- Encode as a polynomial, like the contrasts of R factors.
The contrasts are new to me, I'm more used to the one-hot encoding. So there will be an additional command-line flag, named "--encoding", taking several values:
- "none";
- "one-hot";
- "yes" (defaults to one-hot).