[
https://issues.apache.org/jira/browse/MAHOUT-1249?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13975130#comment-13975130
]
jian wang commented on MAHOUT-1249:
-----------------------------------
OK thanks for your sharing.
Originally, the paper of Netflix provides their stopping criterion:
"
The stopping criterion we use is based on the observed RMSE on
the probe dataset. After one round of updating both U and M , if the difference
between the observed RMSEs on the probe dataset is less than 1 bps^2, the
iteration stops and we use the obtained U, M to make final predictions on the
test dataset
"
So my original thought on this issue accords with the paper's criterion. And we
may provide parameter interface to specify their customized RMSE threshold and
we could combine it with the numIterations in the stopping criterion.
I shall look to see the spark implementation of ALS to see if this has already
been considered or not.
> Build tools around mahout to check the training error of factorization and
> automatically detect convergence
> -----------------------------------------------------------------------------------------------------------
>
> Key: MAHOUT-1249
> URL: https://issues.apache.org/jira/browse/MAHOUT-1249
> Project: Mahout
> Issue Type: Improvement
> Components: Collaborative Filtering
> Affects Versions: 1.0
> Reporter: Saikat Kanjilal
> Fix For: 1.0
>
>
> The goal of this task is to check the training error of the factorization
> during the computation to make it automatically detect convergence. The goal
> is not to have to specify the number of iterations as a parameter needed for
> convergence.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
