[
https://issues.apache.org/jira/browse/MAHOUT-1058?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13434936#comment-13434936
]
Sean Owen commented on MAHOUT-1058:
-----------------------------------
I may be misinterpreting from reading the patch quickly, but, I had imagined
that the change was nearly all contained to CandidateItemStrategy and code that
uses that.
IDRescorer already filters items being considered for recommendation, in the
recommend methods. I thought the intent was a different kind of filtering,
filtering items considered earlier in a stage of candidate selection. For
example, in item-based recs, I find a pool of items A that the current user
likes. I find a pool of users B that also like stuff in A. Finally I find a
pool of items C that are liked by users in B.
The current IDRescorer application already filters C. I thought the idea was to
accept a different instance to filter A (?). Or somehow B (?).
I guess I am saying that the patch looks redundant. You filter items that would
be filtered in the following step already, before estimation.
> IDRescorers should also be able to filter candidate items
> ---------------------------------------------------------
>
> Key: MAHOUT-1058
> URL: https://issues.apache.org/jira/browse/MAHOUT-1058
> Project: Mahout
> Issue Type: New Feature
> Components: Collaborative Filtering
> Affects Versions: 0.8
> Reporter: Sebastian Schelter
> Assignee: Sebastian Schelter
> Attachments: MAHOUT-1058.patch
>
>
> Currently when a recommendation for a particular user is computed, at first
> the candidate items are determined via a CandidateItemsStrategy, then
> predicitions for all these items are computed and afterwards the IDRescorer
> can filter the items that might be recommended.
> For performance reasons, it should also be possible to apply custom filtering
> before the estimation is done. This cannot be done in the
> CandidateItemsStrategy, as the filtering logic might depend on some
> request/or user-specific criteria that are not known in the recommender
> application.
> It would be possible to implement such a logic in IDRescorer.isFiltered(),
> this is however bad for performance reasons, as the very costly estimation
> has already happen then.
> I propose to add a method to IDRescorer that can filter the candidate items
> right before estimation starts.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
