Handling datasets with multiple targets? about linfa

[bytesnake]

yes playing around with this right now, because it is a prerequisite for #66 I will open a PR once I have something useful

[bytesnake]

work started in #88, I think the basic mechanism is in place and I will finish the PR in the next days

[relf]

A suggestion: now that we have several targets, we could have target_names as well.

[bytesnake]

Hey,
this wasn't considered for now but should be as multi-target prediction is very common.

The targets matrix can be 1, 2 or three dimensional in three different cases:

• 1dim - 1 target per N observations
• 2dim - M targets per N observations
• 3dim - L probabilities for M targets per N observations

here is L the number of labels, M the number of targets and N the number of points. Because ndarray supports three dimensional arrays we should switch the Dataset definition to

pub type Dataset<D, T> = DatasetBase<ArrayBase<OwnedRepr<D>, Ix2>, ArrayBase<OwnedRepr<T>, Ix3>>;

and then update the whole machinery in linfa/dataset/.

[bytesnake]

I added a point to #70 for this feature

[relf]

Ok thanks. I was too conservative and not aware of the third case (do you have an example in mind?)

I added a point to #70 for this feature

Does it mean you plan to work on that in the coming weeks?

[bytesnake]

finally merged #88, can you give feedback whether everything works as expected?

[relf]

Ok, I will. I will resume my work on linnerud dataset. FYI, I am working on a port of PlsRegression which is tested against that dataset. If it goes well, maybe it could land in linfa? Where? A new member linfa-cross-decomposition or ...? What do you think?

[bytesnake]

sounds good, we should add CCA at a later point as well. If you want to implement the SVD version, you can use TruncatedSvd from ndarray-linalg

[bytesnake]

multi-target datasets were introduced #88 and a sample dataset added in #89. I also added a note in #70 regarding target naming