[
https://issues.apache.org/jira/browse/SPARK-9011?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Shivam Verma reopened SPARK-9011:
---------------------------------
I did some more experiments. It is really a bug because
pyspark.ml.tuning.CrossValidator seems to accept outputs of only certain
classifiers. So it is the question of making a design choice: either ensuring
consistency across classifier outputs in Spark.ML or making the
BinaryClassificationEvaluator generic enough.
I have appropriately modified the description above and I am reopening the
issue.
> Spark 1.4.0| Spark.ML Classifier Output Formats Inconsistent --> Grid search
> working on LR but not on RF
> --------------------------------------------------------------------------------------------------------
>
> Key: SPARK-9011
> URL: https://issues.apache.org/jira/browse/SPARK-9011
> Project: Spark
> Issue Type: Bug
> Components: ML, MLlib, PySpark
> Affects Versions: 1.4.0
> Environment: Spark 1.4.0 standalone on top of Hadoop 2.3 on single
> node running CentOS
> Reporter: Shivam Verma
> Priority: Critical
> Labels: cross-validation, ml, mllib, pyspark, randomforest,
> tuning
>
> Hi,
> I ran into this bug while using pyspark.ml.tuning.CrossValidator on an RF
> (Random Forest) classifier to classify a small dataset using the
> pyspark.ml.tuning module. (This is a bug because CrossValidator works on LR
> (Logistic Regression) but not on RF)
> Bug:
> There is an issue with how BinaryClassificationEvaluator(self,
> rawPredictionCol="rawPrediction", labelCol="label",
> metricName="areaUnderROC") interprets the 'rawPredict' column - with LR, the
> rawPredictionCol is expected to contain vectors, whereas with RF, the
> prediction column contains doubles.
> Suggested Resolution: Either enable BinaryClassificationEvaluator to work
> with doubles, or let RF output a column rawPredictions containing the
> probability vectors (with probability of 1 assigned to predicted label, and 0
> assigned to the rest).
> Detailed Observation:
> While running grid search on an RF classifier to classify a small dataset
> using the pyspark.ml.tuning module, specifically the ParamGridBuilder and
> CrossValidator classes. I get the following error when I try passing a
> DataFrame of Features-Labels to CrossValidator:
> {noformat}
> Py4JJavaError: An error occurred while calling o1464.evaluate.
> : java.lang.IllegalArgumentException: requirement failed: Column
> rawPrediction must be of type org.apache.spark.mllib.linalg.VectorUDT@1eef
> but was actually DoubleType.
> {noformat}
> I tried the following code, using the dataset given in Spark's CV
> documentation for [cross
> validator|https://spark.apache.org/docs/latest/api/python/pyspark.ml.html#pyspark.ml.tuning.CrossValidator].
> I also pass the DF through a StringIndexer transformation for the RF:
>
> {noformat}
> dataset = sqlContext.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)] *
> 10,["features", "label"])
> stringIndexer = StringIndexer(inputCol="label", outputCol="indexed")
> si_model = stringIndexer.fit(dataset)
> dataset2 = si_model.transform(dataset)
> keep = [dataset2.features, dataset2.indexed]
> dataset3 = dataset2.select(*keep).withColumnRenamed('indexed','label')
> rf =
> RandomForestClassifier(predictionCol="rawPrediction",featuresCol="features",numTrees=5,
> maxDepth=7)
> grid = ParamGridBuilder().addGrid(rf.maxDepth, [4,5,6]).build()
> evaluator = BinaryClassificationEvaluator()
> cv = CrossValidator(estimator=rf, estimatorParamMaps=grid,
> evaluator=evaluator)
> cvModel = cv.fit(dataset3)
> {noformat}
> Note that the above dataset *works* on logistic regression. I have also tried
> a larger dataset with sparse vectors as features (which I was originally
> trying to fit) but received the same error on RF.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
