You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@mahout.apache.org by "Hao Zhong (JIRA)" <ji...@apache.org> on 2017/03/23 01:34:41 UTC
[jira] [Updated] (MAHOUT-1959) BallKMeans.iterativeAssignment can
set wrong weights.
[ https://issues.apache.org/jira/browse/MAHOUT-1959?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Hao Zhong updated MAHOUT-1959:
------------------------------
Status: Patch Available (was: Open)
diff --git a/mr/src/main/java/org/apache/mahout/clustering/streaming/cluster/BallKMeans.java b/mr/src/main/java/org/apache/mahout/clustering/streaming/cluster/BallKMeans.java
index 25806fe..350f189 100644
--- a/mr/src/main/java/org/apache/mahout/clustering/streaming/cluster/BallKMeans.java
+++ b/mr/src/main/java/org/apache/mahout/clustering/streaming/cluster/BallKMeans.java
@@ -436,7 +436,8 @@
}
for (WeightedVector datapoint : datapoints) {
Centroid closestCentroid = (Centroid) centroids.searchFirst(datapoint, false).getValue();
- closestCentroid.setWeight(closestCentroid.getWeight() + datapoint.getWeight());
+ double closestCentroidWeight = centroids.searchFirst(datapoint, false).getWeight();
+ closestCentroid.setWeight(closestCentroidWeight + datapoint.getWeight());
}
}
}
> BallKMeans.iterativeAssignment can set wrong weights.
> -----------------------------------------------------
>
> Key: MAHOUT-1959
> URL: https://issues.apache.org/jira/browse/MAHOUT-1959
> Project: Mahout
> Issue Type: Bug
> Reporter: Hao Zhong
>
> I notice that the BallKMeans.iterativeAssignment method uses the following code to calculate weights:
> {code:title=BallKMeans.java|borderStyle=solid}
> for (WeightedVector datapoint : datapoints) {
> Centroid closestCentroid = (Centroid) centroids.searchFirst(datapoint, false).getValue();
> closestCentroid.setWeight(closestCentroid.getWeight() + datapoint.getWeight());
> }
> {code}
> In MAHOUT-1237, the buggy code is the same way to calculate the weight:
> {code:title=ClusteringUtils.java|borderStyle=solid}
> for (Vector vector : datapoints) {
> Centroid closest = (Centroid) centroids.searchFirst(vector, false).getValue();
> totalCost += closest.getWeight();
> }
> {code}
> The fixed code is as follow:
> {code:title=ClusteringUtils.java|borderStyle=solid}
> for (Vector vector : datapoints) {
> totalCost += centroids.searchFirst(vector, false).getWeight();
> }
> {code}
> I am not quite sure whether BallKMeans.iterativeAssignment sets the right weights. Please check it.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)