[R] unsupervised dimensionality reduction retaining the attributes

[Rainer Tischler]

Dear all,

I'm looking for an unsupervised dimensionality reduction method for large-scale data sets (> 40.000 attributes) which retains the original attribute vectors (i.e. "unsupervised feature selection"). Until now I have only found several unsupervised dimensionality reduction methods which use derived features (transforming instead of selecting features) and some very simple variance filters (e.g. fold change filtering).
Is anybody aware of a more advanced method, which can score feature subsets instead of just single features and is fast enough for high-dimensional data sets (like the greedy CFS-method for supervised feature subset selection, but in an unsupervised setting)? If there is no R-package for this, I would also be glad if someone could point me to a good research paper about this topic. I have no labels for my data; hence, supervised feature selection is not an option for me.

many thanks in advance,
Rainer