ML Tuning: model selection and hyperparameter tuning

\[ \newcommand{\R}{\mathbb{R}} \newcommand{\E}{\mathbb{E}} \newcommand{\x}{\mathbf{x}} \newcommand{\y}{\mathbf{y}} \newcommand{\wv}{\mathbf{w}} \newcommand{\av}{\mathbf{\alpha}} \newcommand{\bv}{\mathbf{b}} \newcommand{\N}{\mathbb{N}} \newcommand{\id}{\mathbf{I}} \newcommand{\ind}{\mathbf{1}} \newcommand{\0}{\mathbf{0}} \newcommand{\unit}{\mathbf{e}} \newcommand{\one}{\mathbf{1}} \newcommand{\zero}{\mathbf{0}} \]

This section describes how to use MLlib’s tooling for tuning ML algorithms and Pipelines. Built-in Cross-Validation and other tooling allow users to optimize hyperparameters in algorithms and Pipelines.

Table of contents

Model selection (a.k.a. hyperparameter tuning)

An important task in ML is model selection, or using data to find the best model or parameters for a given task. This is also called tuning. Tuning may be done for individual Estimators such as LogisticRegression, or for entire Pipelines which include multiple algorithms, featurization, and other steps. Users can tune an entire Pipeline at once, rather than tuning each element in the Pipeline separately.

MLlib supports model selection using tools such as CrossValidator and TrainValidationSplit. These tools require the following items:

At a high level, these model selection tools work as follows:

The Evaluator can be a RegressionEvaluator for regression problems, a BinaryClassificationEvaluator for binary data, or a MulticlassClassificationEvaluator for multiclass problems. The default metric used to choose the best ParamMap can be overridden by the setMetricName method in each of these evaluators.

To help construct the parameter grid, users can use the ParamGridBuilder utility.

Cross-Validation

CrossValidator begins by splitting the dataset into a set of folds which are used as separate training and test datasets. E.g., with $k=3$ folds, CrossValidator will generate 3 (training, test) dataset pairs, each of which uses 2/3 of the data for training and 1/3 for testing. To evaluate a particular ParamMap, CrossValidator computes the average evaluation metric for the 3 Models produced by fitting the Estimator on the 3 different (training, test) dataset pairs.

After identifying the best ParamMap, CrossValidator finally re-fits the Estimator using the best ParamMap and the entire dataset.

Example: model selection via cross-validation

The following example demonstrates using CrossValidator to select from a grid of parameters.

Note that cross-validation over a grid of parameters is expensive. E.g., in the example below, the parameter grid has 3 values for hashingTF.numFeatures and 2 values for lr.regParam, and CrossValidator uses 2 folds. This multiplies out to $(3 \times 2) \times 2 = 12$ different models being trained. In realistic settings, it can be common to try many more parameters and use more folds ($k=3$ and $k=10$ are common). In other words, using CrossValidator can be very expensive. However, it is also a well-established method for choosing parameters which is more statistically sound than heuristic hand-tuning.

import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.classification.LogisticRegression
import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator
import org.apache.spark.ml.feature.{HashingTF, Tokenizer}
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.ml.tuning.{CrossValidator, ParamGridBuilder}
import org.apache.spark.sql.Row

// Prepare training data from a list of (id, text, label) tuples.
val training = spark.createDataFrame(Seq(
  (0L, "a b c d e spark", 1.0),
  (1L, "b d", 0.0),
  (2L, "spark f g h", 1.0),
  (3L, "hadoop mapreduce", 0.0),
  (4L, "b spark who", 1.0),
  (5L, "g d a y", 0.0),
  (6L, "spark fly", 1.0),
  (7L, "was mapreduce", 0.0),
  (8L, "e spark program", 1.0),
  (9L, "a e c l", 0.0),
  (10L, "spark compile", 1.0),
  (11L, "hadoop software", 0.0)
)).toDF("id", "text", "label")

// Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
val tokenizer = new Tokenizer()
  .setInputCol("text")
  .setOutputCol("words")
val hashingTF = new HashingTF()
  .setInputCol(tokenizer.getOutputCol)
  .setOutputCol("features")
val lr = new LogisticRegression()
  .setMaxIter(10)
val pipeline = new Pipeline()
  .setStages(Array(tokenizer, hashingTF, lr))

// We use a ParamGridBuilder to construct a grid of parameters to search over.
// With 3 values for hashingTF.numFeatures and 2 values for lr.regParam,
// this grid will have 3 x 2 = 6 parameter settings for CrossValidator to choose from.
val paramGrid = new ParamGridBuilder()
  .addGrid(hashingTF.numFeatures, Array(10, 100, 1000))
  .addGrid(lr.regParam, Array(0.1, 0.01))
  .build()

// We now treat the Pipeline as an Estimator, wrapping it in a CrossValidator instance.
// This will allow us to jointly choose parameters for all Pipeline stages.
// A CrossValidator requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
// Note that the evaluator here is a BinaryClassificationEvaluator and its default metric
// is areaUnderROC.
val cv = new CrossValidator()
  .setEstimator(pipeline)
  .setEvaluator(new BinaryClassificationEvaluator)
  .setEstimatorParamMaps(paramGrid)
  .setNumFolds(2)  // Use 3+ in practice

// Run cross-validation, and choose the best set of parameters.
val cvModel = cv.fit(training)

// Prepare test documents, which are unlabeled (id, text) tuples.
val test = spark.createDataFrame(Seq(
  (4L, "spark i j k"),
  (5L, "l m n"),
  (6L, "mapreduce spark"),
  (7L, "apache hadoop")
)).toDF("id", "text")

// Make predictions on test documents. cvModel uses the best model found (lrModel).
cvModel.transform(test)
  .select("id", "text", "probability", "prediction")
  .collect()
  .foreach { case Row(id: Long, text: String, prob: Vector, prediction: Double) =>
    println(s"($id, $text) --> prob=$prob, prediction=$prediction")
  }
Find full example code at "examples/src/main/scala/org/apache/spark/examples/ml/ModelSelectionViaCrossValidationExample.scala" in the Spark repo.
import java.util.Arrays;

import org.apache.spark.ml.Pipeline;
import org.apache.spark.ml.PipelineStage;
import org.apache.spark.ml.classification.LogisticRegression;
import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator;
import org.apache.spark.ml.feature.HashingTF;
import org.apache.spark.ml.feature.Tokenizer;
import org.apache.spark.ml.param.ParamMap;
import org.apache.spark.ml.tuning.CrossValidator;
import org.apache.spark.ml.tuning.CrossValidatorModel;
import org.apache.spark.ml.tuning.ParamGridBuilder;
import org.apache.spark.sql.Dataset;
import org.apache.spark.sql.Row;

// Prepare training documents, which are labeled.
Dataset<Row> training = spark.createDataFrame(Arrays.asList(
  new JavaLabeledDocument(0L, "a b c d e spark", 1.0),
  new JavaLabeledDocument(1L, "b d", 0.0),
  new JavaLabeledDocument(2L,"spark f g h", 1.0),
  new JavaLabeledDocument(3L, "hadoop mapreduce", 0.0),
  new JavaLabeledDocument(4L, "b spark who", 1.0),
  new JavaLabeledDocument(5L, "g d a y", 0.0),
  new JavaLabeledDocument(6L, "spark fly", 1.0),
  new JavaLabeledDocument(7L, "was mapreduce", 0.0),
  new JavaLabeledDocument(8L, "e spark program", 1.0),
  new JavaLabeledDocument(9L, "a e c l", 0.0),
  new JavaLabeledDocument(10L, "spark compile", 1.0),
  new JavaLabeledDocument(11L, "hadoop software", 0.0)
), JavaLabeledDocument.class);

// Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
Tokenizer tokenizer = new Tokenizer()
  .setInputCol("text")
  .setOutputCol("words");
HashingTF hashingTF = new HashingTF()
  .setNumFeatures(1000)
  .setInputCol(tokenizer.getOutputCol())
  .setOutputCol("features");
LogisticRegression lr = new LogisticRegression()
  .setMaxIter(10)
  .setRegParam(0.01);
Pipeline pipeline = new Pipeline()
  .setStages(new PipelineStage[] {tokenizer, hashingTF, lr});

// We use a ParamGridBuilder to construct a grid of parameters to search over.
// With 3 values for hashingTF.numFeatures and 2 values for lr.regParam,
// this grid will have 3 x 2 = 6 parameter settings for CrossValidator to choose from.
ParamMap[] paramGrid = new ParamGridBuilder()
  .addGrid(hashingTF.numFeatures(), new int[] {10, 100, 1000})
  .addGrid(lr.regParam(), new double[] {0.1, 0.01})
  .build();

// We now treat the Pipeline as an Estimator, wrapping it in a CrossValidator instance.
// This will allow us to jointly choose parameters for all Pipeline stages.
// A CrossValidator requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
// Note that the evaluator here is a BinaryClassificationEvaluator and its default metric
// is areaUnderROC.
CrossValidator cv = new CrossValidator()
  .setEstimator(pipeline)
  .setEvaluator(new BinaryClassificationEvaluator())
  .setEstimatorParamMaps(paramGrid).setNumFolds(2);  // Use 3+ in practice

// Run cross-validation, and choose the best set of parameters.
CrossValidatorModel cvModel = cv.fit(training);

// Prepare test documents, which are unlabeled.
Dataset<Row> test = spark.createDataFrame(Arrays.asList(
  new JavaDocument(4L, "spark i j k"),
  new JavaDocument(5L, "l m n"),
  new JavaDocument(6L, "mapreduce spark"),
  new JavaDocument(7L, "apache hadoop")
), JavaDocument.class);

// Make predictions on test documents. cvModel uses the best model found (lrModel).
Dataset<Row> predictions = cvModel.transform(test);
for (Row r : predictions.select("id", "text", "probability", "prediction").collectAsList()) {
  System.out.println("(" + r.get(0) + ", " + r.get(1) + ") --> prob=" + r.get(2)
    + ", prediction=" + r.get(3));
}
Find full example code at "examples/src/main/java/org/apache/spark/examples/ml/JavaModelSelectionViaCrossValidationExample.java" in the Spark repo.
from pyspark.ml import Pipeline
from pyspark.ml.classification import LogisticRegression
from pyspark.ml.evaluation import BinaryClassificationEvaluator
from pyspark.ml.feature import HashingTF, Tokenizer
from pyspark.ml.tuning import CrossValidator, ParamGridBuilder

# Prepare training documents, which are labeled.
training = spark.createDataFrame([
    (0, "a b c d e spark", 1.0),
    (1, "b d", 0.0),
    (2, "spark f g h", 1.0),
    (3, "hadoop mapreduce", 0.0),
    (4, "b spark who", 1.0),
    (5, "g d a y", 0.0),
    (6, "spark fly", 1.0),
    (7, "was mapreduce", 0.0),
    (8, "e spark program", 1.0),
    (9, "a e c l", 0.0),
    (10, "spark compile", 1.0),
    (11, "hadoop software", 0.0)
], ["id", "text", "label"])

# Configure an ML pipeline, which consists of tree stages: tokenizer, hashingTF, and lr.
tokenizer = Tokenizer(inputCol="text", outputCol="words")
hashingTF = HashingTF(inputCol=tokenizer.getOutputCol(), outputCol="features")
lr = LogisticRegression(maxIter=10)
pipeline = Pipeline(stages=[tokenizer, hashingTF, lr])

# We now treat the Pipeline as an Estimator, wrapping it in a CrossValidator instance.
# This will allow us to jointly choose parameters for all Pipeline stages.
# A CrossValidator requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
# We use a ParamGridBuilder to construct a grid of parameters to search over.
# With 3 values for hashingTF.numFeatures and 2 values for lr.regParam,
# this grid will have 3 x 2 = 6 parameter settings for CrossValidator to choose from.
paramGrid = ParamGridBuilder() \
    .addGrid(hashingTF.numFeatures, [10, 100, 1000]) \
    .addGrid(lr.regParam, [0.1, 0.01]) \
    .build()

crossval = CrossValidator(estimator=pipeline,
                          estimatorParamMaps=paramGrid,
                          evaluator=BinaryClassificationEvaluator(),
                          numFolds=2)  # use 3+ folds in practice

# Run cross-validation, and choose the best set of parameters.
cvModel = crossval.fit(training)

# Prepare test documents, which are unlabeled.
test = spark.createDataFrame([
    (4L, "spark i j k"),
    (5L, "l m n"),
    (6L, "mapreduce spark"),
    (7L, "apache hadoop")
], ["id", "text"])

# Make predictions on test documents. cvModel uses the best model found (lrModel).
prediction = cvModel.transform(test)
selected = prediction.select("id", "text", "probability", "prediction")
for row in selected.collect():
    print(row)
Find full example code at "examples/src/main/python/ml/cross_validator.py" in the Spark repo.

Train-Validation Split

In addition to CrossValidator Spark also offers TrainValidationSplit for hyper-parameter tuning. TrainValidationSplit only evaluates each combination of parameters once, as opposed to k times in the case of CrossValidator. It is therefore less expensive, but will not produce as reliable results when the training dataset is not sufficiently large.

Unlike CrossValidator, TrainValidationSplit creates a single (training, test) dataset pair. It splits the dataset into these two parts using the trainRatio parameter. For example with $trainRatio=0.75$, TrainValidationSplit will generate a training and test dataset pair where 75% of the data is used for training and 25% for validation.

Like CrossValidator, TrainValidationSplit finally fits the Estimator using the best ParamMap and the entire dataset.

Example: model selection via train validation split

import org.apache.spark.ml.evaluation.RegressionEvaluator
import org.apache.spark.ml.regression.LinearRegression
import org.apache.spark.ml.tuning.{ParamGridBuilder, TrainValidationSplit}

// Prepare training and test data.
val data = spark.read.format("libsvm").load("data/mllib/sample_linear_regression_data.txt")
val Array(training, test) = data.randomSplit(Array(0.9, 0.1), seed = 12345)

val lr = new LinearRegression()

// We use a ParamGridBuilder to construct a grid of parameters to search over.
// TrainValidationSplit will try all combinations of values and determine best model using
// the evaluator.
val paramGrid = new ParamGridBuilder()
  .addGrid(lr.regParam, Array(0.1, 0.01))
  .addGrid(lr.fitIntercept)
  .addGrid(lr.elasticNetParam, Array(0.0, 0.5, 1.0))
  .build()

// In this case the estimator is simply the linear regression.
// A TrainValidationSplit requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
val trainValidationSplit = new TrainValidationSplit()
  .setEstimator(lr)
  .setEvaluator(new RegressionEvaluator)
  .setEstimatorParamMaps(paramGrid)
  // 80% of the data will be used for training and the remaining 20% for validation.
  .setTrainRatio(0.8)

// Run train validation split, and choose the best set of parameters.
val model = trainValidationSplit.fit(training)

// Make predictions on test data. model is the model with combination of parameters
// that performed best.
model.transform(test)
  .select("features", "label", "prediction")
  .show()
Find full example code at "examples/src/main/scala/org/apache/spark/examples/ml/ModelSelectionViaTrainValidationSplitExample.scala" in the Spark repo.
import org.apache.spark.ml.evaluation.RegressionEvaluator;
import org.apache.spark.ml.param.ParamMap;
import org.apache.spark.ml.regression.LinearRegression;
import org.apache.spark.ml.tuning.ParamGridBuilder;
import org.apache.spark.ml.tuning.TrainValidationSplit;
import org.apache.spark.ml.tuning.TrainValidationSplitModel;
import org.apache.spark.sql.Dataset;
import org.apache.spark.sql.Row;

Dataset<Row> data = spark.read().format("libsvm")
  .load("data/mllib/sample_linear_regression_data.txt");

// Prepare training and test data.
Dataset<Row>[] splits = data.randomSplit(new double[] {0.9, 0.1}, 12345);
Dataset<Row> training = splits[0];
Dataset<Row> test = splits[1];

LinearRegression lr = new LinearRegression();

// We use a ParamGridBuilder to construct a grid of parameters to search over.
// TrainValidationSplit will try all combinations of values and determine best model using
// the evaluator.
ParamMap[] paramGrid = new ParamGridBuilder()
  .addGrid(lr.regParam(), new double[] {0.1, 0.01})
  .addGrid(lr.fitIntercept())
  .addGrid(lr.elasticNetParam(), new double[] {0.0, 0.5, 1.0})
  .build();

// In this case the estimator is simply the linear regression.
// A TrainValidationSplit requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
TrainValidationSplit trainValidationSplit = new TrainValidationSplit()
  .setEstimator(lr)
  .setEvaluator(new RegressionEvaluator())
  .setEstimatorParamMaps(paramGrid)
  .setTrainRatio(0.8);  // 80% for training and the remaining 20% for validation

// Run train validation split, and choose the best set of parameters.
TrainValidationSplitModel model = trainValidationSplit.fit(training);

// Make predictions on test data. model is the model with combination of parameters
// that performed best.
model.transform(test)
  .select("features", "label", "prediction")
  .show();
Find full example code at "examples/src/main/java/org/apache/spark/examples/ml/JavaModelSelectionViaTrainValidationSplitExample.java" in the Spark repo.
from pyspark.ml.evaluation import RegressionEvaluator
from pyspark.ml.regression import LinearRegression
from pyspark.ml.tuning import ParamGridBuilder, TrainValidationSplit

# Prepare training and test data.
data = spark.read.format("libsvm")\
    .load("data/mllib/sample_linear_regression_data.txt")
train, test = data.randomSplit([0.7, 0.3])
lr = LinearRegression(maxIter=10, regParam=0.1)

# We use a ParamGridBuilder to construct a grid of parameters to search over.
# TrainValidationSplit will try all combinations of values and determine best model using
# the evaluator.
paramGrid = ParamGridBuilder()\
    .addGrid(lr.regParam, [0.1, 0.01]) \
    .addGrid(lr.elasticNetParam, [0.0, 0.5, 1.0])\
    .build()

# In this case the estimator is simply the linear regression.
# A TrainValidationSplit requires an Estimator, a set of Estimator ParamMaps, and an Evaluator.
tvs = TrainValidationSplit(estimator=lr,
                           estimatorParamMaps=paramGrid,
                           evaluator=RegressionEvaluator(),
                           # 80% of the data will be used for training, 20% for validation.
                           trainRatio=0.8)

# Run TrainValidationSplit, and choose the best set of parameters.
model = tvs.fit(train)
# Make predictions on test data. model is the model with combination of parameters
# that performed best.
prediction = model.transform(test)
for row in prediction.take(5):
    print(row)
Find full example code at "examples/src/main/python/ml/train_validation_split.py" in the Spark repo.