Hey Deb,
If your goal is to solve the subproblems in ALS, exploring sparsity
doesn't give you much benefit because the data is small and dense.
Porting either ECOS's or PDCO's implementation but using dense
representation should be sufficient. Feel free to open a JIRA and we
can move our
I looked further and realized that ECOS used a mex file while PDCO is using
pure Matlab code. So the out-of-box runtime comparison is not fair.
I am trying to generate PDCO C port. Like ECOS, PDCO also makes use of
sparse support from Tim Davis.
Thanks.
Deb
Hi Xiangrui,
I did some out-of-box comparisons with ECOS and PDCO from SOL.
Both of them seems to be running at par but I will do more detailed
analysis.
I used pdco's testQP randomized problem generation. pdcotestQP(m, n) means
m constraints and n variables
For ECOS runtime reference here is
Hi Xiangrui,
Could you please point to the IPM solver that you have positive results
with ? I was planning to compare with CVX, KNITRO from Professor Nocedal
and MOSEK probably...I don't have CPLEX license so I won't be able to do
that comparison...
My experiments so far tells me that ADMM based
Hi Deb,
KNITRO and MOSEK are both commercial. If you are looking for
open-source ones, you can take a look at PDCO from SOL:
http://web.stanford.edu/group/SOL/software/pdco/
Each subproblem is really just a small QP. ADMM is designed for the
cases when data is distributively stored or the
For explicit feedback, ALS uses only observed ratings for computation.
So XtXs are not the same. -Xiangrui
On Tue, Jun 10, 2014 at 8:58 PM, Debasish Das debasish.da...@gmail.com wrote:
Sorry last one went out by mistake:
Is not for users (0 to numUsers), fullXtX is same ? In the ALS
You idea is close to what implicit feedback does. You can check the
paper, which is short and concise. In the ALS setting, all subproblems
are independent in each iteration. This is part of the reason why ALS
is scalable. If you have some global constraints that make the
subproblems no longer
Sorry last one went out by mistake:
Is not for users (0 to numUsers), fullXtX is same ? In the ALS formulation
this is W^TW or H^TH which should be same for all the users ? Why we are
reading userXtX(index) and adding it to fullXtX in the loop over all
numUsers ?
// Solve the least-squares
I don't quite understand why putting linear constraints can promote
orthogonality. For the interfaces, if the subproblem is determined by
Y^T Y and Y^T b for each iteration, then the least squares solver, the
non-negative least squares solver, or your convex solver is simply a
function
(A, b) -
Hi Xiangrui,
It's not the linear constraint, It is quadratic inequality with slack,
first order taylor approximation of off diagonal cross terms and a cyclic
coordinate descent, which we think will yield orthogonalityIt's still
under works...
Also we want to put a L1 constraint as set of
Hi,
We are adding a constrained ALS solver in Spark to solve matrix
factorization use-cases which needs additional constraints (bounds,
equality, inequality, quadratic constraints)
We are using a native version of a primal dual SOCP solver due to its small
memory footprint and sparse ccs matrix
Hi Deb,
Why do you want to make those methods public? If you only need to
replace the solver for subproblems. You can try to make the solver
pluggable. Now it supports least squares and non-negative least
squares. You can define an interface for the subproblem solvers and
maintain the IPM solver
Hi Xiangrui,
For orthogonality properties in the factors we need a constraint solver
other than the usuals (l1, upper and lower bounds, l2 etc)
The interface of constraint solver is standard and I can add it in mllib
optimization
But I am not sure how will I call the gpl licensed ipm solver
13 matches
Mail list logo
