flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Trevor Grant (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-2259) Support training Estimators using a (train, validation, test) split of the available data
Date Fri, 12 Feb 2016 14:02:18 GMT

    [ https://issues.apache.org/jira/browse/FLINK-2259?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15144606#comment-15144606
] 

Trevor Grant commented on FLINK-2259:
-------------------------------------

In the spirit of sci-kit, and just data stuff in general:

The plan of attack should be to make a 2-way random split (with split options), then make
2 types of wrappers:

One for a train-test-validate style,

Another for k-folding, which would require some more functionality but would more 'grown-up'
style. 

So basically I plan on making a function, and then another function just to wrap the first-
which may seem silly, but the anticipation of k-folds is my motivation. 

> Support training Estimators using a (train, validation, test) split of the available
data
> -----------------------------------------------------------------------------------------
>
>                 Key: FLINK-2259
>                 URL: https://issues.apache.org/jira/browse/FLINK-2259
>             Project: Flink
>          Issue Type: New Feature
>          Components: Machine Learning Library
>            Reporter: Theodore Vasiloudis
>            Assignee: Trevor Grant
>            Priority: Minor
>              Labels: ML
>
> When there is an abundance of data available, a good way to train models is to split
the available data into 3 parts: Train, Validation and Test.
> We use the Train data to train the model, the Validation part is used to estimate the
test error and select hyperparameters, and the Test is used to evaluate the performance of
the model, and assess its generalization [1]
> This is a common approach when training Artificial Neural Networks, and a good strategy
to choose in data-rich environments. Therefore we should have some support of this data-analysis
process in our Estimators.
> [1] Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. The elements of statistical
learning. Vol. 1. Springer, Berlin: Springer series in statistics, 2001.



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

Mime
View raw message