Return type changes depending on user input; this is surprising and likely to break code

[st--]

The data returned by ScatterWidget depends on user input, specifically whether the user only ever added points of one, or of more colors:

• one color only ("regression style"): widget.data_as_X_y returns X as an [N,1] shaped array of x coordinates and y as an [N] shaped array (vector) of y coordinates
• two+ colors ("classification style"): widget.data_as_X_y returns X as an [N,2] shaped array of x and y coordinates and y as an [N] shaped array of labels

It's great that the widget supports both regression and classification, but this makes it hard to build notebooks on top of it. Generally, I would want to handle both very differently. It would be great to be able to enforce at widget creation time that I
(a) only want regression (and then don't provide the radio buttons for classes) or
(b) only want classification (in which case always return X as [N,2] array and y as [N] array with labels, even if there is only a single class).

[koaning]

Yeah, I agree that the API is a bit wonky and implicit here, which is not ideal.

Originally the barwidget was considered to be a way to have a 1-1 x/y mapping, which could cover a lot of regression tasks, but I agree that a more explicit scatter widget could help make things better. I will consider this point in a few days, can't really say when exactly, but a small revamp of this library is long overdue. I want more consistent styles for UI across elements and topics like this could be taken into account too.