This is pretty cool!

A small feature request: I uploaded data with a combination of continuous features and discrete features (which happened to be coded as numbers—2 or 1—with negative numbers used to indicate various cases was missing). I'm pretty sure the model interpreted the discrete features as continuous. I think it'd be good to add a detector for numerical fields that take on a small number of values, and convert those to binary features.

Likewise, non-linear models. I know my data has a lot of internal structure and so a tree, neural network, etc. would capture a lot more of the variance.