[jira] [Resolved] (SPARK-27867) RegressionEvaluator cache lastest RegressionMetrics to avoid duplicated computation

["zhengruifeng (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SPARK-27867?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

zhengruifeng resolved SPARK-27867.
----------------------------------
    Resolution: Not A Problem

> RegressionEvaluator cache lastest RegressionMetrics to avoid duplicated computation
> -----------------------------------------------------------------------------------
>
>                 Key: SPARK-27867
>                 URL: https://issues.apache.org/jira/browse/SPARK-27867
>             Project: Spark
>          Issue Type: Improvement
>          Components: ML
>    Affects Versions: 3.0.0
>            Reporter: zhengruifeng
>            Priority: Major
>
> In most cases, given a model, we have to obtain multi metrics of it.
> For examples, a regression model, we may need to obtain the R2, MAE and MSE.
> However, current design of `Evaluator` do not support computing multi metrics at once.
> In practice, we usually use RegressionEvaluator like this:
> {code:java}
> val evaluator = new RegressionEvaluator()
> val r2 = evaluator.setMetricName("r2").evaluate(df)
> val mae = evaluator.setMetricName("mae").evaluate(df)
> val mse = evaluator.setMetricName("mse").evaluate(df){code}
>  
> However, current impl of RegressionEvaluator needs one pass of the whole input dataset to compute one metric. So, above example needs 3 passes.
> This can be optimized since in \{RegressionMetrics}  all metrics can be computed at once.
> If we cache the lastest inputs, and then if the next evaluate call keep the inputs (except the metricName), then we can directly obtain the metric from the internal intermediate summary.
>  



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org