[ https://issues.apache.org/jira/browse/SPARK-4039?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14336177#comment-14336177 ]
Derrick Burns commented on SPARK-4039: -------------------------------------- Support sparse data in the clusterer is a bad idea. But so is converting sparse data directly to a dense representation for any data of significant dimension. So, what do you do? The HashingTF is one approach to reduce the dimension of the data, but it is not a good one. Collisions can lead to dramatic overestimates. Instead, randoming indexing should be used. Randoming Indexing (http://en.wikipedia.org/wiki/Random_indexing), per the Johnson-Lindenstrauss Lemma, guarantees that the embedding in a lower dimension space preserves certain distance measures. See https://github.com/derrickburns/generalized-kmeans-clustering for an implementation. > KMeans support sparse cluster centers > ------------------------------------- > > Key: SPARK-4039 > URL: https://issues.apache.org/jira/browse/SPARK-4039 > Project: Spark > Issue Type: Improvement > Components: MLlib > Affects Versions: 1.1.0 > Reporter: Antoine Amend > Labels: clustering > > When the number of features is not known, it might be quite helpful to create > sparse vectors using HashingTF.transform. KMeans transforms centers vectors > to dense vectors > (https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala#L307), > therefore leading to OutOfMemory (even with small k). > Any way to keep vectors sparse ? -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org