[
https://issues.apache.org/jira/browse/SPARK-2401?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14621607#comment-14621607
]
Gang Bai edited comment on SPARK-2401 at 7/10/15 5:04 AM:
----------------------------------------------------------
Yes, this should be based on a basic abstraction of AdaBoost or additive
models, and use Pipelines API. BTW, I implemented the AdaBoost.MH algo,
experimentally and in a separate repo. It was based on Spark/MLLib 1.1, so it's
not with Pipelines API. It's in the repo
https://github.com/BaiGang/spark_multiboost, and also I made some presentations
about this work https://github.com/BaiGang/slides/tree/master/spark_multiboost.
I am willing to improve it, ideally make it base on SPARK-1546 as part of
mllib and use Pipelines API.
was (Author: baigang):
Yes, this should be based on a basic abstraction of AdaBoost or additive
models, and use Pipelines API. BTW, I implemented the AdaBoost.MH algo,
experimentally and in a separate repo. It was based on Spark/MLLib 1.1, so it's
not with Pipelines API. It's in the repo
https://github.com/BaiGang/spark_multiboost, and also I made some presentations
about this work https://github.com/BaiGang/slides/tree/master/spark_multiboost.
I am willing to improve it, ideally make it base on SPARK-1546 and use
Pipelines API and part of mllib.
> AdaBoost.MH, a multi-class multi-label classifier
> -------------------------------------------------
>
> Key: SPARK-2401
> URL: https://issues.apache.org/jira/browse/SPARK-2401
> Project: Spark
> Issue Type: New Feature
> Components: MLlib
> Reporter: Gang Bai
> Priority: Trivial
>
> Multi-class multi-label classifiers are very useful in web page profiling,
> audience segmentation etc. The goal of a multi-class multi-label classifier
> is to tag a sample data point with a subset of labels from a finite,
> pre-specified set, e.g. tagging a visitor with a set of interests. Given a
> set of L labels, a data point can be tagged with one of the 2^L possible
> subsets. The main challenges in training a multi-class multi-label classifier
> are the exponentially large label space.
> This JIRA is created to track the effort of solving the training problem of
> multi-class, multi-label classifiers by implementing AdaBoost.MH on Apache
> Spark. It will not be an easy task. I will start from a basic DecisionStump
> weak learner and a simple Hamming tree resembling DecisionStumps into a meta
> weak learner, and the iterative boosting procedure. I will be reusing modules
> of Alexander Ulanov's multi-class and multi-label metrics evaluation and
> Manish Amde's decision tree/boosting/ensemble implementations.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
