[jira] [Commented] (MATH-1374) KMeansPlusPlusClusterer unable to converge having repeatable points in input dataset

["Artem Barger (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MATH-1374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15311864#comment-15311864 ] 

Artem Barger commented on MATH-1374:
------------------------------------

Reporting this bug here, since I'm not 100% sure what is the desirable or correct solution to fix it, for example {{matlab}} and {{octave}} implementation of {{kmeans}} dealing with such situation, by randomly selecting repeated points into new clusters. Also I'm not sure whenever this is a real life scenario one should take care off. I've found it accidentally running some tests on my data.

> KMeansPlusPlusClusterer unable to converge having repeatable points in input dataset
> ------------------------------------------------------------------------------------
>
>                 Key: MATH-1374
>                 URL: https://issues.apache.org/jira/browse/MATH-1374
>             Project: Commons Math
>          Issue Type: Bug
>            Reporter: Artem Barger
>
> If the input list size of {{Clusterable}} is greater than parameter {{k}} while has less unique points than {{k}}, the algorithm will fail to converge, tested w/ different EmptyClusterStrategy options, here is the example of default one: 
> {code}
>    @Test
>     public void testNumberOfRequestedClustersSameAsInputSize() {
>         final RandomVectorGenerator rng = new UncorrelatedRandomVectorGenerator(10,
>                 new GaussianRandomGenerator(RandomSource.create(RandomSource.MT)));
>         List<DoublePoint> points = new ArrayList<>();
>         for (int i = 0; i < 10; i++) {
>             final DoublePoint point = new DoublePoint(rng.nextVector());
>             for (int j = 0; j < 3; j++) {
>                 points.add(point);
>             }
>         }
>         final KMeansPlusPlusClusterer<DoublePoint> clusterer = new KMeansPlusPlusClusterer<>(12);
>         clusterer.cluster(points);
>     }
> {code}



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