[
https://issues.apache.org/jira/browse/SPARK-18948?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15784216#comment-15784216
]
Joseph K. Bradley commented on SPARK-18948:
-------------------------------------------
Thanks [~danilo.ascione] for suggesting this. A few initial comments:
* We are not accepting new features in the RDD-based API (spark.mllib), but
only in the DataFrame-based API (spark.ml). If you'd like to get this in, it
will need to be rewritten for the DataFrame-based API.
* (minor) Please don't set the Shepherd field; committers will use it to track
releases.
I haven't been able to check out the PR yet, but if broader discussions are
being brought up there, then let's discuss the issues on this JIRA first before
further implementation work.
Thanks!
> Add Mean Percentile Rank metric for ranking algorithms
> ------------------------------------------------------
>
> Key: SPARK-18948
> URL: https://issues.apache.org/jira/browse/SPARK-18948
> Project: Spark
> Issue Type: New Feature
> Components: MLlib
> Reporter: Danilo Ascione
>
> Add the Mean Percentile Rank (MPR) metric for ranking algorithms, as
> described in the paper :
> Hu, Y., Y. Koren, and C. Volinsky. “Collaborative Filtering for Implicit
> Feedback Datasets.” In 2008 Eighth IEEE International Conference on Data
> Mining, 263–72, 2008. doi:10.1109/ICDM.2008.22.
> (http://yifanhu.net/PUB/cf.pdf) (NB: MPR is called "Expected percentile rank"
> in the paper)
> The ALS algorithm for implicit feedback in Spark ML is based on the same
> paper.
> Spark ML lacks an implementation of an appropriate metric for implicit
> feedback, so the MPR metric can fulfill this use case.
> This implementation add the metric to the RankingMetrics class under
> org.apache.spark.mllib.evaluation (SPARK-3568), and it uses the same input
> (prediction and label pairs).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
