[
https://issues.apache.org/jira/browse/SPARK-9005?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Feynman Liang updated SPARK-9005:
---------------------------------
Description: {{RegressionMetrics}} currently computes explainedVariance
using {{summary.variance(1)}} (variance of the residuals) where the [Wikipedia
definition|https://en.wikipedia.org/wiki/Fraction_of_variance_unexplained] uses
the residual sum of squares {{math.pow(summary.normL2(1), 2)}}. The two
coincide only when the predictor is unbiased (e.g. an intercept term is
included in a linear model), but this is not always the case. We should change
to be consistent. (was: {{RegressionMetrics}} currently computes
explainedVariance using {{summary.variance(1)}} (variance of the residuals)
where the [Wikipedia
definition|https://en.wikipedia.org/wiki/Fraction_of_variance_unexplained] uses
the residual sum of squares {{math.pow(summary.normL2(1), 2)}}. The two
coincide only when the predictor is unbiased (e.g. an intercept term is
included in a linear model), but this is not always the case. We should change
to be consistent.
The computation for r2 is also currently incorrect. Multiplying by
{{summary.count - 1}} appears to be trying to compute an adjusted r2, but the
lack of a DoF adjustment in the numerator makes the computation inconsistent
with [Wikipedia's
definition|https://en.wikipedia.org/wiki/Coefficient_of_determination]. Since
{{RegresionMetrics}} is not given the number of regression variables, we should
modify and explicitly document that this computes unadjusted R2.)
> RegressionMetrics computing incorrect explainedVariance and r2
> --------------------------------------------------------------
>
> Key: SPARK-9005
> URL: https://issues.apache.org/jira/browse/SPARK-9005
> Project: Spark
> Issue Type: Bug
> Components: MLlib
> Reporter: Feynman Liang
>
> {{RegressionMetrics}} currently computes explainedVariance using
> {{summary.variance(1)}} (variance of the residuals) where the [Wikipedia
> definition|https://en.wikipedia.org/wiki/Fraction_of_variance_unexplained]
> uses the residual sum of squares {{math.pow(summary.normL2(1), 2)}}. The two
> coincide only when the predictor is unbiased (e.g. an intercept term is
> included in a linear model), but this is not always the case. We should
> change to be consistent.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
