[SPARK-16832] CrossValidator and TrainValidationSplit are not random without seed

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Won't Fix
• Affects Version/s: 2.0.0
• Component/s: ML, PySpark
• Assignee: Unassigned
• Reporter: Max Moroz
• Created: 01/Aug/16 08:47
• Updated: 30/Aug/16 10:36
• Resolved: 30/Aug/16 10:35

Description

Repeatedly running CrossValidator or TrainValidationSplit without an explicit seed parameter does not change results. It is supposed to be seeded with a random seed, but it seems to be instead seeded with some constant. (If seed is explicitly provided, the two classes behave as expected.)

dataset = spark.createDataFrame(
  [(Vectors.dense([0.0]), 0.0),
   (Vectors.dense([0.4]), 1.0),
   (Vectors.dense([0.5]), 0.0),
   (Vectors.dense([0.6]), 1.0),
   (Vectors.dense([1.0]), 1.0)] * 1000,
  ["features", "label"]).cache()

paramGrid = pyspark.ml.tuning.ParamGridBuilder().build()
tvs = pyspark.ml.tuning.TrainValidationSplit(estimator=pyspark.ml.regression.LinearRegression(), 
                           estimatorParamMaps=paramGrid,
                           evaluator=pyspark.ml.evaluation.RegressionEvaluator(),
                           trainRatio=0.8)
model = tvs.fit(train)
print(model.validationMetrics)

for folds in (3, 5, 10):
  cv = pyspark.ml.tuning.CrossValidator(estimator=pyspark.ml.regression.LinearRegression(), 
                                      estimatorParamMaps=paramGrid, 
                                      evaluator=pyspark.ml.evaluation.RegressionEvaluator(),
                                      numFolds=folds
                                     )
  cvModel = cv.fit(dataset)
  print(folds, cvModel.avgMetrics)

This code produces identical results upon repeated calls.

Issue Links

• relates to: SPARK-17311 Standardize Python-Java MLlib API to accept optional long seeds in all cases
• links to: [Github] Pull Request #14524 (srowen)