Manifold learners are not transformers #2887
Comments
|
I bump into this issue today when I wanted to make a pipeline with a spectral embedding. |
|
same here - i think it would be good to add a transform method to these. they should act no differently from PCA imo. |
|
This issue comes from my (initially confused) question on Stackoverflow. It would be great to be able to project unseen points into the MDS space, but after thinking about it, I don't think the MDS algorithm can actually do this! Can it? I think it iteratively "moves points around" in the projected space. It doesn't actually represent the transformation. I suppose there is a possibility to hack it by find a few nearest neighbours of the unseen point in the original space, and sending it to a weighted average of their transforms or something like that. But there would be no justification in MDS theory for that, I think. |
|
i see where this is coming from now. indeed for many of these operators one can create a transform for the input data set, but it's not possible to apply that transformation to a different dataset. so i take back what i said earlier about adding the transform bit - i was indeed thinking the fit_transform process, which they already have. |
But what is the algorithm to transform on new data? I am not sure that I |
Unless we can come up with a principled algorithm to transform new data, |
|
I just stumbled upon this -- you can indeed project new points in MDS and other methods. See |
MDSandSpectralEmbeddingare not transformers: they havefit_transform, but notransform. Of these,SpectralEmbeddinginherits fromTransformerMixinwithout good cause. Either it should be removed or the class should get atransformmethod.Ping @GaelVaroquaux, @kuantkid, @NelleV.
The text was updated successfully, but these errors were encountered: