[
https://issues.apache.org/jira/browse/SPARK-8540?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14633883#comment-14633883
]
Joseph K. Bradley commented on SPARK-8540:
------------------------------------------
On the one hand, I agree this could potentially be solved with a good code
example.
On the other hand, it is another cognitive step for users looking to do outlier
detection. Also, I suspect we will eventually want complex algorithms
specialized for outlier/anomaly detection. If we only put complex outlier
detection algorithms under the name "outlier detection," then users may use
those unnecessarily complex algorithms by default. E.g., I suspect this
happens a lot in sklearn, where the only one explicitly under "outlier
detection" is 1-class SVM, which is surely overkill for many use cases.
> KMeans-based outlier detection
> ------------------------------
>
> Key: SPARK-8540
> URL: https://issues.apache.org/jira/browse/SPARK-8540
> Project: Spark
> Issue Type: Sub-task
> Components: ML
> Reporter: Joseph K. Bradley
> Original Estimate: 336h
> Remaining Estimate: 336h
>
> Proposal for K-Means-based outlier detection:
> * Cluster data using K-Means
> * Provide prediction/filtering functionality which returns outliers/anomalies
> ** This can take some threshold parameter which specifies either (a) how far
> off a point needs to be to be considered an outlier or (b) how many outliers
> should be returned.
> Note this will require a bit of API design, which should probably be posted
> and discussed on this JIRA before implementation.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
