[
https://issues.apache.org/jira/browse/SPARK-19957?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
yuhao yang updated SPARK-19957:
-------------------------------
Description:
when users set the initialization mode to "random", KMeans in ML and MLlib has
inconsistent behavior for multiple runs:
MLlib will basically use new Random for each run.
ML Kmeans however will use the default random seed, which is
{code}this.getClass.getName.hashCode.toLong{code}, and keep using the same
number among multiple fitting.
I would expect the "random" initialization mode to be literally random.
There're different solutions with different scope of impact. Adjusting the
hasSeed trait may have a broader impact(but maybe worth discussion). We can
always just set random default seed in KMeans.
Appreciate your feedback.
was:
when users set the initialization mode to "random", KMeans in ML and MLlib has
inconsistent behavior for multiple runs:
MLlib will basically use new Random for each run.
ML Kmeans however will use the default random seed, which is
{code}this.getClass.getName.hashCode.toLong{code}, and keep using the same
number among multiple fitting.
I would expect the "random" initialization mode to be literally random.
There're different solutions with different scope of impact. Adjusting the
hasSeed trait may have a broader impact. We can always just set random default
seed in KMeans.
Appreciate your feedback.
> Inconsist KMeans initialization mode behavior between ML and MLlib
> ------------------------------------------------------------------
>
> Key: SPARK-19957
> URL: https://issues.apache.org/jira/browse/SPARK-19957
> Project: Spark
> Issue Type: Bug
> Components: ML
> Affects Versions: 2.1.0
> Reporter: yuhao yang
> Priority: Minor
>
> when users set the initialization mode to "random", KMeans in ML and MLlib
> has inconsistent behavior for multiple runs:
> MLlib will basically use new Random for each run.
> ML Kmeans however will use the default random seed, which is
> {code}this.getClass.getName.hashCode.toLong{code}, and keep using the same
> number among multiple fitting.
> I would expect the "random" initialization mode to be literally random.
> There're different solutions with different scope of impact. Adjusting the
> hasSeed trait may have a broader impact(but maybe worth discussion). We can
> always just set random default seed in KMeans.
> Appreciate your feedback.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
