You are viewing a plain text version of this content. The canonical link for it is here.

Posted to issues@commons.apache.org by "Gilles (JIRA)" <ji...@apache.org> on 2016/05/06 22:00:14 UTC

[jira] [Updated] (MATH-1315) MultiKMeansPlusPlusClusterer buggy for alternative evaluators

     [ https://issues.apache.org/jira/browse/MATH-1315?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Gilles updated MATH-1315:
-------------------------
    Attachment: MATH-1356.patch

Do you agree that the attached patch would solve this issue?

> MultiKMeansPlusPlusClusterer buggy for alternative evaluators
> -------------------------------------------------------------
>
>                 Key: MATH-1315
>                 URL: https://issues.apache.org/jira/browse/MATH-1315
>             Project: Commons Math
>          Issue Type: Bug
>    Affects Versions: 3.5
>            Reporter: Thorsten Schäfer
>         Attachments: MATH-1356.patch
>
>
> I just looked into the source code for the MultiKMeansPlusPlusClusterer and realized that it would return null in case of an alternative evaluator that favors bigger values instead of smaller ones:
> The basic idea of the clustering method is that we perform n clusterings and choose the best result. The decision what's the best result is performed by the evaluator, which by default assumes smaller values are better. 
> According to the documentation, we can also provide a different evaluator, which for instance would decide that bigger values are better, but given we initialize the best value with Double.POSITIVE_INFINITY in method MultiKMeansPlusPlusClusterer.cluster(Collection<T>), we would never find a "better" result and thus always return null. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)