[
https://issues.apache.org/jira/browse/MAHOUT-672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13023332#comment-13023332
]
Jake Mannix commented on MAHOUT-672:
------------------------------------
Yes, a SquaredMatrix and a PlusIdentityMultipleMatrix (? ugly name) would be
enough, if composable properly.
We might need two variants, sadly. Maybe we should migrate VectorIterable to
some other abstract base class (get rid of interface, for previously discussed
interface/abstract class reasons), and give it a better name. Maybe that would
make it easier to *not* have two variants? It would be a class which just has
the times(Vector) and times(Matrix) methods, and that's almost it?
(numRows/numCols too, I guess).
As for LanczosSolver, please check out the patch for MAHOUT-319. The api for
solve is most likely changing anyways. And I'm in favor of just *removing*
timesSquared() and isSymmetric, not marking deprecated. Still pre-1.0-days,
folks! :)
> Implementation of Conjugate Gradient for solving large linear systems
> ---------------------------------------------------------------------
>
> Key: MAHOUT-672
> URL: https://issues.apache.org/jira/browse/MAHOUT-672
> Project: Mahout
> Issue Type: New Feature
> Components: Math
> Affects Versions: 0.5
> Reporter: Jonathan Traupman
> Priority: Minor
> Attachments: 0001-MAHOUT-672-LSMR-iterative-linear-solver.patch,
> 0001-MAHOUT-672-LSMR-iterative-linear-solver.patch, MAHOUT-672.patch,
> MAHOUT-672.patch
>
> Original Estimate: 48h
> Remaining Estimate: 48h
>
> This patch contains an implementation of conjugate gradient, an iterative
> algorithm for solving large linear systems. In particular, it is well suited
> for large sparse systems where a traditional QR or Cholesky decomposition is
> infeasible. Conjugate gradient only works for matrices that are square,
> symmetric, and positive definite (basically the same types where Cholesky
> decomposition is applicable). Systems like these commonly occur in statistics
> and machine learning problems (e.g. regression).
> Both a standard (in memory) solver and a distributed hadoop-based solver
> (basically the standard solver run using a DistributedRowMatrix a la
> DistributedLanczosSolver) are included.
> There is already a version of this algorithm in taste package, but it doesn't
> operate on standard mahout matrix/vector objects, nor does it implement a
> distributed version. I believe this implementation will be more generically
> useful to the community than the specialized one in taste.
> This implementation solves the following types of systems:
> Ax = b, where A is square, symmetric, and positive definite
> A'Ax = b where A is arbitrary but A'A is positive definite. Directly solving
> this system is more efficient than computing A'A explicitly then solving.
> (A + lambda * I)x = b and (A'A + lambda * I)x = b, for systems where A or A'A
> is singular and/or not full rank. This occurs commonly if A is large and
> sparse. Solving a system of this form is used, for example, in ridge
> regression.
> In addition to the normal conjugate gradient solver, this implementation also
> handles preconditioning, and has a sample Jacobi preconditioner included as
> an example. More work will be needed to build more advanced preconditioners
> if desired.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
