[jira] [Created] (MAHOUT-1941) Use the existing regression infrastructure to implement Logistic Regression using Samsara

Discussion:

Saikat Kanjilal (JIRA)

2017-02-12 19:34:42 UTC

Permalink

Saikat Kanjilal created MAHOUT-1941:
---------------------------------------

Summary: Use the existing regression infrastructure to implement Logistic Regression using Samsara
Key: MAHOUT-1941
URL: https://issues.apache.org/jira/browse/MAHOUT-1941
Project: Mahout
Issue Type: New Feature
Components: Classification
Affects Versions: 1.0.0
Reporter: Saikat Kanjilal
Fix For: 1.0.0

The goal is to reuse this chunk of mahout infrastructure to implement logistic regression: https://github.com/apache/mahout/tree/master/math-scala/src/main/scala/org/apache/mahout/math/algorithms/regression

I need this for a work related POC and will get started by extending the RegressorModel.scala, will also put forward a design proposal on the dev mailing list.

--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Saikat Kanjilal (JIRA)

2017-02-12 19:37:41 UTC

Permalink

[ https://issues.apache.org/jira/browse/MAHOUT-1941?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15862930#comment-15862930 ]

Saikat Kanjilal commented on MAHOUT-1941:
-----------------------------------------

[~rawkintrevo] Since I need this for a work related POC I went ahead and created this JIRA to implement logistic regression based on the new algorithms object model that you created, I will send out a design on this JIRA as well as dev list first to review before beginning implementation, the goal is to implement logistic regression using multinomial/binary logistic regression

Post by Saikat Kanjilal (JIRA)
Use the existing regression infrastructure to implement Logistic Regression using Samsara
-----------------------------------------------------------------------------------------
Key: MAHOUT-1941
URL: https://issues.apache.org/jira/browse/MAHOUT-1941
Project: Mahout
Issue Type: New Feature
Components: Classification
Affects Versions: 1.0.0
Reporter: Saikat Kanjilal
Labels: features
Fix For: 1.0.0
The goal is to reuse this chunk of mahout infrastructure to implement logistic regression: https://github.com/apache/mahout/tree/master/math-scala/src/main/scala/org/apache/mahout/math/algorithms/regression
I need this for a work related POC and will get started by extending the RegressorModel.scala, will also put forward a design proposal on the dev mailing list.

--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Saikat Kanjilal (JIRA)

2017-03-05 00:18:33 UTC

Permalink

[ https://issues.apache.org/jira/browse/MAHOUT-1941?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15895982#comment-15895982 ]

Saikat Kanjilal commented on MAHOUT-1941:
-----------------------------------------

This issue will be solved when GLM is fully implemented at scale as that will include logistic regression

--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Trevor Grant (JIRA)

2017-03-05 02:15:33 UTC

Permalink

[ https://issues.apache.org/jira/browse/MAHOUT-1941?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15896030#comment-15896030 ]

Trevor Grant commented on MAHOUT-1941:
--------------------------------------

Should have replied here, not dev. Nice work- but a tip is to extend LinearRegressorFitter / Model.

Even if you need to override for calculating standard error (since the Betas may not be normal) you'll avoid a lot of boilerplate code on the fitter tests.. You can just override as needed (e.g. you can override calculate stand error).

It's also possible there are things in LinearRegressorModel (and Fitter) that really belong in OLS, and don't generalize to all linear models. I don't claim to be infallible :)

--
This message was sent by Atlassian JIRA
(v6.3.15#6346)