[
https://issues.apache.org/jira/browse/SPARK-2426?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14176358#comment-14176358
]
Debasish Das commented on SPARK-2426:
-------------------------------------
[~mengxr] I thought more on it and one of the reason we choose ADMM because
QuadraticMinimizer is not designed to be a local algorithm....
If it runs on Spark Master it will take a RDD...If it runs on Spark worker, it
will take a H and c from x'Hx + c'x along with proximal operators...
I will update the API and show some POCs that how this meta algorithm will add
LBFGS/Truncated Newton as a core solver for x-solve for scalable version of
matrix factorization where we don't want to create the H matrix explicitly
ever...
Truncated Newton is a better choice for the constraints we want to support...I
am working on a variant of TRON and linear CG that's in breeze for the scalable
version..Those are the building blocks I need...
I am sure some of the code will move to Breeze. Proximal will definitely move
to Breeze but QuadraticMinimizer will be refactored. It will really help if you
can open up a PR on the new ALS design you have and we can work on it...
> Quadratic Minimization for MLlib ALS
> ------------------------------------
>
> Key: SPARK-2426
> URL: https://issues.apache.org/jira/browse/SPARK-2426
> Project: Spark
> Issue Type: New Feature
> Components: MLlib
> Affects Versions: 1.0.0
> Reporter: Debasish Das
> Assignee: Debasish Das
> Original Estimate: 504h
> Remaining Estimate: 504h
>
> Current ALS supports least squares and nonnegative least squares.
> I presented ADMM and IPM based Quadratic Minimization solvers to be used for
> the following ALS problems:
> 1. ALS with bounds
> 2. ALS with L1 regularization
> 3. ALS with Equality constraint and bounds
> Initial runtime comparisons are presented at Spark Summit.
> http://spark-summit.org/2014/talk/quadratic-programing-solver-for-non-negative-matrix-factorization-with-spark
> Based on Xiangrui's feedback I am currently comparing the ADMM based
> Quadratic Minimization solvers with IPM based QpSolvers and the default
> ALS/NNLS. I will keep updating the runtime comparison results.
> For integration the detailed plan is as follows:
> 1. Add QuadraticMinimizer and Proximal algorithms in mllib.optimization
> 2. Integrate QuadraticMinimizer in mllib ALS
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
