[
https://issues.apache.org/jira/browse/SPARK-2426?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14175167#comment-14175167
]
Debasish Das commented on SPARK-2426:
-------------------------------------
1. [~mengxr] Our legal was clear that Stanford and Verizon copyright should
show up on the COPYRIGHT.txt file...I saw other company's copyrights and I did
not think it will be a big issue...
2. For the new interface, we have two more requirements: Convex loss function
(supporting huber loss / hinge loss etc) and no explicit AtA construction since
once we start scaling to 10000 factors for LSA then AtA construction will start
to choke...Can I work on your branch ?
https://github.com/mengxr/spark-als/blob/master/src/main/scala/org/apache/spark/ml/SimpleALS.scala
3. I agree to refactor the core solver including NNLS to breeze. That was the
initial plan but since we wanted to test out the features in our internal
datasets, integrating with mllib was faster. I am testing NNLS's CG
implementation since as soon as explicit AtA construction is taken out, we need
to rely on CG in-place of direct solvers...But I will refactor the solver out
to breeze and that will take the copyright msgs to breeze as well.
4. Let me add the Matlab scripts and point to the repository. ECOS and MOSEK
will need Matlab to run. PDCO and Proximal variants will run fine on Octave. I
am not sure if MOSEK is supported on Octave.
> 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: [email protected]
For additional commands, e-mail: [email protected]
