From 0b8995168f02bb55afb0a5b7dbdb941c3c89cb4c Mon Sep 17 00:00:00 2001 From: Liang-Chi Hsieh Date: Sat, 20 Jun 2015 13:01:59 -0700 Subject: [PATCH] [SPARK-8468] [ML] Take the negative of some metrics in RegressionEvaluator to get correct cross validation JIRA: https://issues.apache.org/jira/browse/SPARK-8468 Author: Liang-Chi Hsieh Closes #6905 from viirya/cv_min and squashes the following commits: 930d3db [Liang-Chi Hsieh] Fix python unit test and add document. d632135 [Liang-Chi Hsieh] Merge remote-tracking branch 'upstream/master' into cv_min 16e3b2c [Liang-Chi Hsieh] Take the negative instead of reciprocal. c3dd8d9 [Liang-Chi Hsieh] For comments. b5f52c1 [Liang-Chi Hsieh] Add param to CrossValidator for choosing whether to maximize evaulation value. --- .../ml/evaluation/RegressionEvaluator.scala | 10 ++++-- .../org/apache/spark/ml/param/params.scala | 2 +- .../evaluation/RegressionEvaluatorSuite.scala | 4 +-- .../spark/ml/tuning/CrossValidatorSuite.scala | 35 +++++++++++++++++-- python/pyspark/ml/evaluation.py | 8 +++-- 5 files changed, 48 insertions(+), 11 deletions(-) diff --git a/mllib/src/main/scala/org/apache/spark/ml/evaluation/RegressionEvaluator.scala b/mllib/src/main/scala/org/apache/spark/ml/evaluation/RegressionEvaluator.scala index 8670e9679d055..01c000b47514c 100644 --- a/mllib/src/main/scala/org/apache/spark/ml/evaluation/RegressionEvaluator.scala +++ b/mllib/src/main/scala/org/apache/spark/ml/evaluation/RegressionEvaluator.scala @@ -37,6 +37,10 @@ final class RegressionEvaluator(override val uid: String) /** * param for metric name in evaluation (supports `"rmse"` (default), `"mse"`, `"r2"`, and `"mae"`) + * + * Because we will maximize evaluation value (ref: `CrossValidator`), + * when we evaluate a metric that is needed to minimize (e.g., `"rmse"`, `"mse"`, `"mae"`), + * we take and output the negative of this metric. * @group param */ val metricName: Param[String] = { @@ -70,13 +74,13 @@ final class RegressionEvaluator(override val uid: String) val metrics = new RegressionMetrics(predictionAndLabels) val metric = $(metricName) match { case "rmse" => - metrics.rootMeanSquaredError + -metrics.rootMeanSquaredError case "mse" => - metrics.meanSquaredError + -metrics.meanSquaredError case "r2" => metrics.r2 case "mae" => - metrics.meanAbsoluteError + -metrics.meanAbsoluteError } metric } diff --git a/mllib/src/main/scala/org/apache/spark/ml/param/params.scala b/mllib/src/main/scala/org/apache/spark/ml/param/params.scala index 15ebad8838a2a..50c0d855066f8 100644 --- a/mllib/src/main/scala/org/apache/spark/ml/param/params.scala +++ b/mllib/src/main/scala/org/apache/spark/ml/param/params.scala @@ -297,7 +297,7 @@ class DoubleArrayParam(parent: Params, name: String, doc: String, isValid: Array /** * :: Experimental :: - * A param amd its value. + * A param and its value. */ @Experimental case class ParamPair[T](param: Param[T], value: T) { diff --git a/mllib/src/test/scala/org/apache/spark/ml/evaluation/RegressionEvaluatorSuite.scala b/mllib/src/test/scala/org/apache/spark/ml/evaluation/RegressionEvaluatorSuite.scala index aa722da323935..5b203784559e2 100644 --- a/mllib/src/test/scala/org/apache/spark/ml/evaluation/RegressionEvaluatorSuite.scala +++ b/mllib/src/test/scala/org/apache/spark/ml/evaluation/RegressionEvaluatorSuite.scala @@ -63,7 +63,7 @@ class RegressionEvaluatorSuite extends SparkFunSuite with MLlibTestSparkContext // default = rmse val evaluator = new RegressionEvaluator() - assert(evaluator.evaluate(predictions) ~== 0.1019382 absTol 0.001) + assert(evaluator.evaluate(predictions) ~== -0.1019382 absTol 0.001) // r2 score evaluator.setMetricName("r2") @@ -71,6 +71,6 @@ class RegressionEvaluatorSuite extends SparkFunSuite with MLlibTestSparkContext // mae evaluator.setMetricName("mae") - assert(evaluator.evaluate(predictions) ~== 0.08036075 absTol 0.001) + assert(evaluator.evaluate(predictions) ~== -0.08036075 absTol 0.001) } } diff --git a/mllib/src/test/scala/org/apache/spark/ml/tuning/CrossValidatorSuite.scala b/mllib/src/test/scala/org/apache/spark/ml/tuning/CrossValidatorSuite.scala index 36af4b34a9e40..db64511a76055 100644 --- a/mllib/src/test/scala/org/apache/spark/ml/tuning/CrossValidatorSuite.scala +++ b/mllib/src/test/scala/org/apache/spark/ml/tuning/CrossValidatorSuite.scala @@ -20,11 +20,12 @@ package org.apache.spark.ml.tuning import org.apache.spark.SparkFunSuite import org.apache.spark.ml.{Estimator, Model} import org.apache.spark.ml.classification.LogisticRegression -import org.apache.spark.ml.evaluation.{BinaryClassificationEvaluator, Evaluator} +import org.apache.spark.ml.evaluation.{BinaryClassificationEvaluator, Evaluator, RegressionEvaluator} import org.apache.spark.ml.param.ParamMap import org.apache.spark.ml.param.shared.HasInputCol +import org.apache.spark.ml.regression.LinearRegression import org.apache.spark.mllib.classification.LogisticRegressionSuite.generateLogisticInput -import org.apache.spark.mllib.util.MLlibTestSparkContext +import org.apache.spark.mllib.util.{LinearDataGenerator, MLlibTestSparkContext} import org.apache.spark.sql.{DataFrame, SQLContext} import org.apache.spark.sql.types.StructType @@ -58,6 +59,36 @@ class CrossValidatorSuite extends SparkFunSuite with MLlibTestSparkContext { assert(cvModel.avgMetrics.length === lrParamMaps.length) } + test("cross validation with linear regression") { + val dataset = sqlContext.createDataFrame( + sc.parallelize(LinearDataGenerator.generateLinearInput( + 6.3, Array(4.7, 7.2), Array(0.9, -1.3), Array(0.7, 1.2), 100, 42, 0.1), 2)) + + val trainer = new LinearRegression + val lrParamMaps = new ParamGridBuilder() + .addGrid(trainer.regParam, Array(1000.0, 0.001)) + .addGrid(trainer.maxIter, Array(0, 10)) + .build() + val eval = new RegressionEvaluator() + val cv = new CrossValidator() + .setEstimator(trainer) + .setEstimatorParamMaps(lrParamMaps) + .setEvaluator(eval) + .setNumFolds(3) + val cvModel = cv.fit(dataset) + val parent = cvModel.bestModel.parent.asInstanceOf[LinearRegression] + assert(parent.getRegParam === 0.001) + assert(parent.getMaxIter === 10) + assert(cvModel.avgMetrics.length === lrParamMaps.length) + + eval.setMetricName("r2") + val cvModel2 = cv.fit(dataset) + val parent2 = cvModel2.bestModel.parent.asInstanceOf[LinearRegression] + assert(parent2.getRegParam === 0.001) + assert(parent2.getMaxIter === 10) + assert(cvModel2.avgMetrics.length === lrParamMaps.length) + } + test("validateParams should check estimatorParamMaps") { import CrossValidatorSuite._ diff --git a/python/pyspark/ml/evaluation.py b/python/pyspark/ml/evaluation.py index d8ddb78c6d639..595593a7f2cde 100644 --- a/python/pyspark/ml/evaluation.py +++ b/python/pyspark/ml/evaluation.py @@ -160,13 +160,15 @@ class RegressionEvaluator(JavaEvaluator, HasLabelCol, HasPredictionCol): ... >>> evaluator = RegressionEvaluator(predictionCol="raw") >>> evaluator.evaluate(dataset) - 2.842... + -2.842... >>> evaluator.evaluate(dataset, {evaluator.metricName: "r2"}) 0.993... >>> evaluator.evaluate(dataset, {evaluator.metricName: "mae"}) - 2.649... + -2.649... """ - # a placeholder to make it appear in the generated doc + # Because we will maximize evaluation value (ref: `CrossValidator`), + # when we evaluate a metric that is needed to minimize (e.g., `"rmse"`, `"mse"`, `"mae"`), + # we take and output the negative of this metric. metricName = Param(Params._dummy(), "metricName", "metric name in evaluation (mse|rmse|r2|mae)")