[jira] [Assigned] (SPARK-10188) Pyspark CrossValidator with RMSE selects incorrect model

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

     [ https://issues.apache.org/jira/browse/SPARK-10188?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Apache Spark reassigned SPARK-10188:
------------------------------------

    Assignee:     (was: Apache Spark)

> Pyspark CrossValidator with RMSE selects incorrect model
> --------------------------------------------------------
>
>                 Key: SPARK-10188
>                 URL: https://issues.apache.org/jira/browse/SPARK-10188
>             Project: Spark
>          Issue Type: Bug
>          Components: PySpark
>    Affects Versions: 1.5.0
>            Reporter: Noel Smith
>
> Pyspark {{CrossValidator}} is giving incorrect results when selecting estimators using RMSE as an evaluation metric.
> In the example below, it should be selecting the {{LogisticRegression}} estimator with zero regularization as that gives the most accurate result, but instead it selects the one with the largest.
> Probably related to: SPARK-10097
> {code}
> from pyspark.ml.evaluation import RegressionEvaluator
> from pyspark.ml.regression import LinearRegression
> from pyspark.ml.tuning import ParamGridBuilder, CrossValidator, CrossValidatorModel
> from pyspark.ml.feature import Binarizer
> from pyspark.mllib.linalg import Vectors
> from pyspark.sql import SQLContext
> sqlContext = SQLContext(sc)
> # Label = 2 * feature
> train = sqlContext.createDataFrame([
>     (Vectors.dense([10.0]), 20.0), 
>     (Vectors.dense([100.0]), 200.0), 
>     (Vectors.dense([1000.0]), 2000.0)] * 10,
>     ["features", "label"])
> test = sqlContext.createDataFrame([
>     (Vectors.dense([1000.0]),)],  
>     ["features"])
> # Expected prediction 2000.0
> print LinearRegression(regParam=0.0).fit(train).transform(test).collect() # Predicts 2000.0 (perfect)
> print LinearRegression(regParam=100.0).fit(train).transform(test).collect() # Predicts 1869.31
> print LinearRegression(regParam=1000000.0).fit(train).transform(test).collect() # 741.08 (worst)
> # Cross-validation
> lr = LinearRegression()
> rmse_eval = RegressionEvaluator()
> grid = (ParamGridBuilder()
>     .addGrid( lr.regParam, [0.0, 100.0, 1000000.0] )
>     .build())
> cv = CrossValidator(estimator=lr, estimatorParamMaps=grid, evaluator=rmse_eval)
> cv_model = cv.fit(train)
> cv_model.bestModel.transform(test).collect() # Predicts 741.08 (i.e. worst model selected)
> {code}
> Once workaround for users would be to add a wrapper around the selected evaluator to invert the metric:
> {code}
> class InvertedEvaluator(Evaluator):
>     def __init__(self, evaluator):
>         super(Evaluator, self).__init__()
>         self.evaluator = evaluator
>     
>     def _evaluate(self, dataset):
>         return -self.evaluator.evaluate(dataset)
>  invertedEvaluator = InvertedEvaluator(RegressionEvaluator(metricName="rmse"))
> {code}



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

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