[
https://issues.apache.org/jira/browse/MAHOUT-1566?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14014747#comment-14014747
]
Dmitriy Lyubimov commented on MAHOUT-1566:
------------------------------------------
primary merit here is an new "showcase" of drm algebra and iron out as many
bugs and issues in preparation for implicit feedback issue. I have had implicit
feedback issue implemented in spark and scala bindings for more than a year
(long before it was added mllib) but it would be hard to address all issues at
once with mahout drm algebra in a bigger method than in a smaller one.
For example, in addition to already listed improvements and fixes in spark
bindings, rmse computation makes it is pretty clear that elementwise logical op
is missing a physical op for identically distributed matrices (which would
produce, again, an identically distributed matrix to operands). It is also
clear that there are probably some bona-fide cases when ABt operator can
produce result identically distributed to A operand if result does not have a
significant size skew (I am still contemplating the criteria of such identical
distribution estimate here).
On a a more supportive argument side, it is obviously false dilemma. Since you
mentioned there's some nonzero merit, and since both methods are of the same
family and can coexist, merit of two is more than merit of either one, not
less. Since drm algebra for the basic one is extremely simple, it has
practically zero maintenance cost.
> Regular ALS factorizer with convergence test.
> ---------------------------------------------
>
> Key: MAHOUT-1566
> URL: https://issues.apache.org/jira/browse/MAHOUT-1566
> Project: Mahout
> Issue Type: Task
> Affects Versions: 0.9
> Reporter: Dmitriy Lyubimov
> Assignee: Dmitriy Lyubimov
> Priority: Trivial
> Fix For: 1.0
>
>
> ALS-related: let's start with unweighed, unregularized implementation.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
