[
https://issues.apache.org/jira/browse/FLINK-3128?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15045013#comment-15045013
]
Fridtjof Sander commented on FLINK-3128:
----------------------------------------
The algorithm is called "Pool Adjacent Violators", which is not a numerical
optimization procedure.
The resulting model of IR is not exactly a function that is parameterized by a
weight vector, but rather a smoothed version of the complete data. Data is
predicted by choosing the nearest label of the smoothed data. Therefore,
previously unseen data that does not reside in the interval of the test data
can not be predicted in a meaningful way.
It looks super inefficient and currently requires that all data can be
processed by one machine.
We haven't really understood for what IR is used for, but it seems to have a
lot of optimization potential, that's why we chose it ;-)
> Add Isotonic Regression To ML Library
> -------------------------------------
>
> Key: FLINK-3128
> URL: https://issues.apache.org/jira/browse/FLINK-3128
> Project: Flink
> Issue Type: New Feature
> Components: Machine Learning Library
> Reporter: Fridtjof Sander
> Assignee: Fridtjof Sander
> Priority: Minor
>
> Isotonic Regression fits a monotonically increasing function (also called
> isotonic function) to a plane of datapoints.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
