fagan2888 / jcur

attempt at CUR decomposition

jCUR

attempt at CUR decomposition

Following the rCUR package as well as the original Matlab (included). This might eventually be useful for use with the Jd database.

Original PNAS paper: http://www.pnas.org/content/106/3/697.full.pdf

If this works and I can make it do something useful, I'll add some other matrix decomposition methods (NNMF, Nuclear Norm regularization, etc) for eventual inclusion into pacman.

Presently, it is completely untested, though the #2 column select technique ties out with the rCUR package in R.

---

The SVD function is not super efficient for large scale problems. In another approximate mx decomp package, the PROPACK lib was a suggested approach in favor of the LAPACK methods. This uses Lanczos bidiagonalization algorithm with partial reorthogonalization (BPRO). Lanczos is simple anyway. Claimed factor of 4 speedup, and 1/2 memory load. http://sun.stanford.edu/~rmunk/PROPACK/paper.ps.gz

R presently uses dgesdd and zgesdd, so as an intermediate step, building the interface for that might help things.

---

The test data set is taken from R, which is in turn taken from the Gene Expression Omnibus database (GSE3443) from the Stanford microarray DB.