[SPARK-24712] TrainValidationSplit ignores label column name and forces to be "label"

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 2.2.0
• Component/s: ML
• Assignee: Unassigned
• Reporter: Pablo J. Villacorta
• Created: 01/Jul/18 18:48
• Updated: 02/Jul/18 11:37
• Resolved: 02/Jul/18 11:37

Description

When a TrainValidationSplit is fit on a Pipeline containing a ML model, the labelCol property of the model is ignored, and the call to fit() will fail unless the labelCol equals "label". As an example, the following pyspark code only works when the variable labelColumn is set to "label"

from pyspark.sql.functions import rand, randn
from pyspark.ml.regression import LinearRegression

labelColumn = "target"  # CHANGE THIS TO "label" AND THE CODE WORKS

df = spark.range(0, 10).select(rand(seed=10).alias("uniform"), randn(seed=27).alias(labelColumn))
vectorAssembler = VectorAssembler().setInputCols(["uniform"]).setOutputCol("features")
lr = LinearRegression().setFeaturesCol("features").setLabelCol(labelColumn)
mypipeline = Pipeline(stages = [vectorAssembler, lr])

paramGrid = ParamGridBuilder()\
.addGrid(lr.regParam, [0.01, 0.1])\
.build()

trainValidationSplit = TrainValidationSplit()\
.setEstimator(mypipeline)\
.setEvaluator(RegressionEvaluator())\
.setEstimatorParamMaps(paramGrid)\
.setTrainRatio(0.8)

trainValidationSplit.fit(df)  # FAIL UNLESS labelColumn IS SET TO "label"