[jira] [Commented] (SPARK-10387) Code generation for decision tree

["DB Tsai (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-10387?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14726392#comment-14726392 ] 

DB Tsai commented on SPARK-10387:
---------------------------------

Here is the current research result from us.

We implemented a prototype of code generation for trees, and here is the implementation of code-gen.
https://github.com/dbtsai/tree/blob/master/macros/src/main/scala/Tree.scala

1) We found the performance of code-gen is 4x to 6x faster than naive binary tree when the # of trees used in GBDT are small. But with around 500x trees, the performance is slightly worse. 

2) We're also benchmarking the flatten trees idea described here, http://tullo.ch/articles/decision-tree-evaluation/

3) Finally, QuickScorer: A Fast Algorithm to Rank Documents with Additive Ensembles of Regression Trees
http://delivery.acm.org/10.1145/2770000/2767733/p73-lucchese.pdf is being implemented, we will benchmark it as well. 

> Code generation for decision tree
> ---------------------------------
>
>                 Key: SPARK-10387
>                 URL: https://issues.apache.org/jira/browse/SPARK-10387
>             Project: Spark
>          Issue Type: New Feature
>          Components: ML
>            Reporter: Xiangrui Meng
>            Assignee: DB Tsai
>
> Provide code generation for decision tree and tree ensembles. Let's first discuss the design and then create new JIRAs for tasks.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org