[
https://issues.apache.org/jira/browse/SPARK-12331?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Imran Younus updated SPARK-12331:
---------------------------------
Description:
The value of R^2 (coefficient of determination) obtained from
LinearRegressionModel is not consistent with R and statsmodels when the
fitIntercept is false i.e., regression through the origin. In this case, both R
and statsmodels use the definition of R^2 given by eq(4') in the following
review paper:
https://online.stat.psu.edu/~ajw13/stat501/SpecialTopics/Reg_thru_origin.pdf
Here is the definition from this paper:
R^2 = \sum(\hat( y)_i^2)/\sum(y_i^2)
The paper also describes why this should be the case. I've double checked that
the value of R^2 from statsmodels and R are consistent with this definition. On
the other hand, scikit-learn doesn't use the above definition. I would
recommend using the above definition in Spark.
was:
The value of R^2 (coefficient of determination) obtained from
LinearRegressionModel is not consistent with R and statsmodels when the
fitIntercept is false i.e., regression through the origin. In this case, both R
and statsmodels use the definition of R^2 given by eq(4') in the following
review paper:
https://online.stat.psu.edu/~ajw13/stat501/SpecialTopics/Reg_thru_origin.pdf
Here is the definition from this paper:
R^2 = \sum(\hat(y)_i^2)/\sum(y_i^2)
The paper also describes why this should be the case. I've double checked that
the value of R^2 from statsmodels and R are consistent with this definition. On
the other hand, scikit-learn doesn't use the above definition. I would
recommend using the above definition in Spark.
> R^2 for regression through the origin
> -------------------------------------
>
> Key: SPARK-12331
> URL: https://issues.apache.org/jira/browse/SPARK-12331
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Reporter: Imran Younus
> Priority: Minor
>
> The value of R^2 (coefficient of determination) obtained from
> LinearRegressionModel is not consistent with R and statsmodels when the
> fitIntercept is false i.e., regression through the origin. In this case, both
> R and statsmodels use the definition of R^2 given by eq(4') in the following
> review paper:
> https://online.stat.psu.edu/~ajw13/stat501/SpecialTopics/Reg_thru_origin.pdf
> Here is the definition from this paper:
> R^2 = \sum(\hat( y)_i^2)/\sum(y_i^2)
> The paper also describes why this should be the case. I've double checked
> that the value of R^2 from statsmodels and R are consistent with this
> definition. On the other hand, scikit-learn doesn't use the above definition.
> I would recommend using the above definition in Spark.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
