Open mateuszbaran opened 3 years ago
I would like to first agree on a framework to use, haven't had time to look at a few (maybe MLJ?)
But first of all I then want to do the simpler ones: Geodesic Regression for example or MDM – but yes in the long run SVM would be nice to have, too, for sure.
I'm not sure we actually need to be exclusive to one particular framework. MLJ integration would be nice for sure but so far I haven't used that package. I've looked at it briefly and I don't really know how to approach that. Somehow we would have to specify that certain data comes from a certain manifold so we'd probably need a whole new family of scientific types: https://github.com/alan-turing-institute/MLJScientificTypes.jl ? I would prefer to figure it out later unless there is a clear need for a framework.
Still, I would like to provide algorithms in this package in a unified way (maybe model, fit, predict), so either we would have to model that ourselves or stich to an existing framework, I think.
fit
and predict
exist at least in StatsBase.jl and MLJModelInterface.jl.
By the way, MDM is essentially kNN for k=1 so PR #1 sort of already covers it.
I've noticed that you (@kellertuer ) have written "SVM". Do you actually mean providing a set of RKHS kernels like in that paper: https://arxiv.org/pdf/1412.0265.pdf ?
EDIT: I just realized that I forgot to mention that I meant SVM from the https://github.com/JuliaManifolds/ManifoldML.jl/blob/master/ideas.md file.