[
https://issues.apache.org/jira/browse/SOLR-15569?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17392231#comment-17392231
]
Alessandro Benedetti commented on SOLR-15569:
---------------------------------------------
To add some information, this is related with this constant:
private static final float NODE_SPLIT_SLACK = 1E-6f; // 0.000001
org/apache/solr/ltr/model/MultipleAdditiveTreesModel.java:125
This little value is added to the threshold transparently to the admin.
[~cpoerschke] , [~diegoceccarelli] [~diegoceccarelli2] [~mnilsson] any idea on
the original reason for that constant to exist?
It seems it has always been there since the original contribution
> MultipleAdditiveTreesModel does not split as expected when the node value
> equals the split threshold
> ----------------------------------------------------------------------------------------------------
>
> Key: SOLR-15569
> URL: https://issues.apache.org/jira/browse/SOLR-15569
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Components: contrib - LTR
> Affects Versions: main (9.0), 8.9
> Reporter: Spyros Kapnissis
> Assignee: Alessandro Benedetti
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> As discussed on the mailing list:
> [https://lists.apache.org/thread.html/r23e10a021445e7bcf26ce578320304bd77165bc0c2136621f7a6225e%40%3Cusers.solr.apache.org%3E]
> , MultipleAdditiveTreesModel does not split as expected when the value of
> the tree node is equal to the threshold, but splits to the left instead of
> the right. This can cause the final score of the model to change
> significantly.
> This issue affects especially categorical or boolean features that take on
> integer values. eg. if a feature can take on 0 or 1 and the left/right split
> is set to 1.0, the right path will never be evaluated.
> I will try to follow up with a test model that showcases the issue.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
