Create Custom Cross Validation in Spark ML

Question

I am new to both Spark and PySpark Data Frames and ML. How can I create a custom cross validation for the ML library. I want for example change the way the training folds are formed, e.g. stratified splits.

This is my current code

numFolds = 10
predictions = []

lr = LogisticRegression()\
     .setFeaturesCol("features")\
     .setLabelCol('label')

# Grid search on LR model
lrparamGrid = ParamGridBuilder()\
     .addGrid(lr.regParam, [0.01, 0.1, 0.5, 1.0, 2.0])\
     .addGrid(lr.elasticNetParam, [0.0, 0.1, 0.5, 0.8, 1.0])\
     .addGrid(lr.maxIter, [5, 10, 20])\
     .build()

pipelineModel = Pipeline(stages=[lr])
evaluator = BinaryClassificationEvaluator()

cv = CrossValidator()\
     .setEstimator(pipelineModel)\
     .setEvaluator(evaluator)\
     .setEstimatorParamMaps(lrparamGrid).setNumFolds(5)

# My own Cross-Validation with stratified splits
for i in range(numFolds):
     # Use Stratified indexOfStratifiedSplits
     trainingData = df[df.ID.isin(indexOfStratifiedSplits[i][0])]
     testingData = df[df.ID.isin(indexOfStratifiedSplits[i][1])]

# Training and Grid Search
cvModel = cv.fit(trainingData)
predictions.append(cvModel.transform(testingData))

I would like to have a Cross-Validation class to be called like this

cv = MyCrossValidator()\
     .setEstimator(pipelineModel)\
     .setEvaluator(evaluator)\
     .setEstimatorParamMaps(lrparamGrid).setNumFolds(5)\
     # Option 1
     .setSplitIndexes(indexOfStratifiedSplits)
     # Option 2
     .setSplitType("Stratified",ColumnName)

I don't know whether the best option is to create a class that extends CrossValidation.fit or Passing Functions to Spark. Either option is challenging for me as newbie, I tried copying GitHub codes but I get tons of errors, specially I don't speak Scala, but this pipeline is faster in Scala API.

Although I have my own functions to split the data in the way I want (sklearn based), I want to use Pipelines, grid search and cv together, such that all permutations are distributed rather than executed in master. That loop with "My own Cross-Validation" only uses part of the cluster nodes as the loop happens in master/driver.

Any Python or Scala API is fine, but preferable Scala.

Thanks

Quenelle

Answer 1

In Python, Sklearn provides you the sklearn.cross_validation.StratifiedKFold function. You can use Sparkit-learn who aims to provide scikit-learn functionality and API on PySpark.

Bawbee

Answer 2

You can also use spark-stratifier if you want a Stratified CrossValidator for PySpark.

Danonorwegian