[
https://issues.apache.org/jira/browse/MAHOUT-305?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12837377#action_12837377
]
Ted Dunning commented on MAHOUT-305:
------------------------------------
My own experience is that all that counts in recommendations is the probability
of click (interest) on a set of recommendations. As such, the best analog is
probably precision at 10 or 20. I don't think that recall at 10 or 20 makes
any sense at all (with a depth limited situation like this, you have given up
on recall and are only looking at precision).
Ankur's suggestion about keeping the most recent 4's and 5's as test data seems
right to me. My only beefs are that you don't need rec...@10 and what to do
with the unrated items. Presumably a new style algorithm could surface items
that the user hadn't thought of, but really likes. In practice, I think that
counting unrated items in the results as misses isn't a big deal in the Netflix
data. In the real world where test data is more scarce, I would count unrated
items as misses in off-line evaluation, but try to run as many alternatives as
possible against live users.
> Combine both cooccurrence-based CF M/R jobs
> -------------------------------------------
>
> Key: MAHOUT-305
> URL: https://issues.apache.org/jira/browse/MAHOUT-305
> Project: Mahout
> Issue Type: Improvement
> Components: Collaborative Filtering
> Affects Versions: 0.2
> Reporter: Sean Owen
> Assignee: Ankur
> Priority: Minor
>
> We have two different but essentially identical MapReduce jobs to make
> recommendations based on item co-occurrence:
> org.apache.mahout.cf.taste.hadoop.{item,cooccurrence}. They ought to be
> merged. Not sure exactly how to approach that but noting this in JIRA, per
> Ankur.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
