Support for metric MDS

[m-muecke]

Implentation: stats::cmdscale

[mb706]

@m-muecke do you know if there is a way to "predict" for new data, i.e. do the same transformation for new points that was done to the training data? This should happen independently of predict data size, i.e. when "predicting" for a dataset, the result should be the same as the rbind() of each row in that dataset predicted independently.

[m-muecke]

I believe there are only efficient solutions when euclidean distances are used, but then the metric MDS is basically equivalent to the PCA which is already implemented.

References:

• De Silva and Tenenbaum (2004). Sparse multidimensional scaling using landmark points.
• Platt (2005). FastMap, MetricMap, and Landmark MDS are all Nystrom Algorithms.

tidymodels use the implementation of the dimRed package, which is only implemented for euclidean distance: https://github.com/gdkrmr/dimRed/blob/master/R/mds.R#L88 and sklearn doesn't seem to have a predict for MDS: https://scikit-learn.org/stable/modules/generated/sklearn.manifold.MDS.html