[
https://issues.apache.org/jira/browse/FLINK-2014?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15484612#comment-15484612
]
Victor Chen commented on FLINK-2014:
------------------------------------
To do LASSO, I'm thinking of implementing Greedy Parallel Coordinate Descent
(http://www.caam.rice.edu/~optimization/disparse/parallel_and_distributed_sparse_optimization.pdf).
Similar to PR#1102 (https://github.com/apache/flink/pull/1102), I'm interested
in performing training with SSP. Would it be too risky to base my alg on SSP
(since the PR hasn't been merged yet)? Are there any "primitives" in Flink I
could use to do non-SSP training (e.g. sync/async training)? I'm planning to
get a minimum viable "product" by the end of this week.
Also, I know Theodore is working on FLINK-2013 (Create GLM framework). Should I
extend the GLM interface from your branch? SInce I plan to finish by the end of
this week and I don't know how much time it'll take, I propose to code a
LASSO-specific algorithm, then within 2 weeks or so, generalize to the GLM
framework.
> Add LASSO regression
> --------------------
>
> Key: FLINK-2014
> URL: https://issues.apache.org/jira/browse/FLINK-2014
> Project: Flink
> Issue Type: New Feature
> Components: Machine Learning Library
> Reporter: Theodore Vasiloudis
> Assignee: Theodore Vasiloudis
> Priority: Minor
>
> LASSO is a linear model that uses L1 regularization
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
