Alessandro Benedetti created SOLR-10449:
-------------------------------------------
Summary: [LTR] Explain summarization improvement for LambdaMART
Key: SOLR-10449
URL: https://issues.apache.org/jira/browse/SOLR-10449
Project: Solr
Issue Type: Improvement
Security Level: Public (Default Security Level. Issues are Public)
Reporter: Alessandro Benedetti
The current explain for the LambdaMART model is quite nice and human readable.
But if you have big ensemble of trees and big trees, it becomes almost
impossible to explain why document has a specific score.
*Scenario*
LambdaMART model with 100 trees, each tree quite tall
A summarized explain in addition, could help.
This could be as advanced as we want :
Simple -> we fetch the features evaluated when scoring, and we return the most
occurring ones
Intermediate -> we return the most occurring features when scoring,
highlighting the positive matches ( go RIGHT for binary features can be more
relevant)
Advanced -> we order first the trees by how much they influenced the final
score and we add this information to the summary weighting differently the
features.
More advanced strategies are welcome, this could really help when explaining
the score of a document running the lambdaMART model.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
