[
https://issues.apache.org/jira/browse/MADLIB-996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15658305#comment-15658305
]
Rahul Iyer commented on MADLIB-996:
-----------------------------------
1. a) Adding new parameters to the CV param list is not a lot of work - I can
add both if interface is defined.
b) Computing a sequence of lambda values is already present in MADlib (called
warmup lambdas in the documentation). It's important to understand that there
are some differences in how the sequences are computed in R compared to MADlib.
Primarily, we have a constant max lambda and don't compute it from the data (as
performed by glmnet).
Further, there are some additional parameters associated warmup process (like
warmup_lambda_no - see docs for details). User need to be aware of these
parameters since the default values might not be the best for their data.
2. CV is used for 2 purposes: model evaluation and model selection. Model
selection does require a grid search (even if the grid is 1 dimensional). 1b
above will make optional the grid input for lambda. Won't resolve the question
for alpha.
Waiting for response to proceed with this ticket.
> Add cross validation support to elastic net
> -------------------------------------------
>
> Key: MADLIB-996
> URL: https://issues.apache.org/jira/browse/MADLIB-996
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Module: Regularized Regression
> Reporter: Frank McQuillan
> Fix For: v1.10
>
>
> http://doc.madlib.net/latest/group__grp__elasticnet.html
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
