Evaluators

A evaluator is a transformation that maps a DataFrame into a metric indicating how good a model is.

evaluator: DataFrame =[evaluate]=> Double

Evaluator is an abstract class with evaluate methods.

evaluate(dataset: DataFrame): Double
evaluate(dataset: DataFrame, paramMap: ParamMap): Double

It employs isLargerBetter method to indicate whether the Double metric should be maximized (true) or minimized (false). It considers a larger value better (true) by default.

isLargerBetter: Boolean = true

The following is a list of some of the available Evaluator implementations:

MulticlassClassificationEvaluator

MulticlassClassificationEvaluator is a concrete Evaluator that expects DataFrame datasets with the following two columns:

  • prediction of DoubleType

  • label of float or double values

BinaryClassificationEvaluator

BinaryClassificationEvaluator is a concrete Evaluator for binary classification that expects datasets (of DataFrame type) with two columns:

  • rawPrediction being DoubleType or VectorUDT.

  • label being NumericType

Note
It can cross-validate models LogisticRegression, RandomForestClassifier et al.

RegressionEvaluator

RegressionEvaluator is a concrete Evaluator for regression that expects datasets (of DataFrame type) with the following two columns:

  • prediction of float or double values

  • label of float or double values

When executed (via evaluate) it prepares a RDD[Double, Double] with (prediction, label) pairs and passes it on to org.apache.spark.mllib.evaluation.RegressionMetrics (from the "old" Spark MLlib).

RegressionEvaluator can evaluate the following metrics:

  • rmse (default; larger is better? no) is the root mean squared error.

  • mse (larger is better? no) is the mean squared error.

  • r2 (larger is better?: yes)

  • mae (larger is better? no) is the mean absolute error.

// prepare a fake input dataset using transformers
import org.apache.spark.ml.feature.Tokenizer
val tok = new Tokenizer().setInputCol("text")

import org.apache.spark.ml.feature.HashingTF
val hashTF = new HashingTF()
  .setInputCol(tok.getOutputCol)  // it reads the output of tok
  .setOutputCol("features")

// Scala trick to chain transform methods
// It's of little to no use since we've got Pipelines
// Just to have it as an alternative
val transform = (tok.transform _).andThen(hashTF.transform _)

val dataset = Seq((0, "hello world", 0.0)).toDF("id", "text", "label")

// we're using Linear Regression algorithm
import org.apache.spark.ml.regression.LinearRegression
val lr = new LinearRegression

import org.apache.spark.ml.Pipeline
val pipeline = new Pipeline().setStages(Array(tok, hashTF, lr))

val model = pipeline.fit(dataset)

// Let's do prediction
// Note that we're using the same dataset as for fitting the model
// Something you'd definitely not be doing in prod
val predictions = model.transform(dataset)

// Now we're ready to evaluate the model
// Evaluator works on datasets with predictions

import org.apache.spark.ml.evaluation.RegressionEvaluator
val regEval = new RegressionEvaluator

// check the available parameters
scala> println(regEval.explainParams)
labelCol: label column name (default: label)
metricName: metric name in evaluation (mse|rmse|r2|mae) (default: rmse)
predictionCol: prediction column name (default: prediction)

scala> regEval.evaluate(predictions)
res0: Double = 0.0

results matching ""

    No results matching ""