[mlpack-svn] [mlpack] add a wrapper for the matrix completion SDP (#374)

[Stephen Tu]

FYI: there is actually a more memory efficient low rank factorization method for the matrix completion SDP. Indeed, given a decision variable X which is m by n, we can factorize it into X = LR^T, where L is m x r and R is n x r. You then apply the same augmented lagrangian technique, and use alternating minimization on the inner optimization (fix R, optimize L. then fix L, optimize R and repeat). See Section 5.3 of http://www.eecs.berkeley.edu/~brecht/papers/07.rfp.lowrank.pdf for more details.

However, there are two reasons I chose to go with the more generic approach. 

  * We can reuse the existing `LRSDP` code.
  * Empirically, I've found that the more generic approach converges faster. That its, the alternating minimization step in the explicit approach outlined above converges quite slowly in my (albeit somewhat limited) experience. 
 
However, it is worth noting that the alternating minimization approach has better theoretical guarantees (see e.g. http://arxiv.org/pdf/1212.0467v1.pdf). 

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/374#issuecomment-68652042
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20150104/bf6e5de1/attachment.html>