spark UnivariateFeatureSelector 源码

spark UnivariateFeatureSelector 代码

文件路径：/mllib/src/main/scala/org/apache/spark/ml/feature/UnivariateFeatureSelector.scala

/*
 * Licensed to the Apache Software Foundation (ASF) under one or more
 * contributor license agreements.  See the NOTICE file distributed with
 * this work for additional information regarding copyright ownership.
 * The ASF licenses this file to You under the Apache License, Version 2.0
 * (the "License"); you may not use this file except in compliance with
 * the License.  You may obtain a copy of the License at
 *
 *    http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package org.apache.spark.ml.feature

import scala.collection.mutable.ArrayBuilder

import org.apache.hadoop.fs.Path

import org.apache.spark.annotation.Since
import org.apache.spark.ml.{Estimator, Model}
import org.apache.spark.ml.attribute.{Attribute, AttributeGroup, NominalAttribute, NumericAttribute}
import org.apache.spark.ml.linalg.{DenseVector, SparseVector, Vector, Vectors, VectorUDT}
import org.apache.spark.ml.param._
import org.apache.spark.ml.param.shared.{HasFeaturesCol, HasLabelCol, HasOutputCol}
import org.apache.spark.ml.stat.{ANOVATest, ChiSquareTest, FValueTest}
import org.apache.spark.ml.util._
import org.apache.spark.sql.{DataFrame, Dataset, SparkSession}
import org.apache.spark.sql.functions.{col, udf}
import org.apache.spark.sql.types.{StructField, StructType}


/**
 * Params for [[UnivariateFeatureSelector]] and [[UnivariateFeatureSelectorModel]].
 */
private[feature] trait UnivariateFeatureSelectorParams extends Params
  with HasFeaturesCol with HasLabelCol with HasOutputCol {

  /**
   * The feature type.
   * Supported options: "categorical", "continuous"
   * @group param
   */
  @Since("3.1.1")
  final val featureType = new Param[String](this, "featureType",
    "Feature type. Supported options: categorical, continuous.",
    ParamValidators.inArray(Array("categorical", "continuous")))

  /** @group getParam */
  @Since("3.1.1")
  def getFeatureType: String = $(featureType)

  /**
   * The label type.
   * Supported options: "categorical", "continuous"
   * @group param
   */
  @Since("3.1.1")
  final val labelType = new Param[String](this, "labelType",
    "Label type. Supported options: categorical, continuous.",
    ParamValidators.inArray(Array("categorical", "continuous")))

  /** @group getParam */
  @Since("3.1.1")
  def getLabelType: String = $(labelType)

  /**
   * The selection mode.
   * Supported options: "numTopFeatures" (default), "percentile", "fpr", "fdr", "fwe"
   * @group param
   */
  @Since("3.1.1")
  final val selectionMode = new Param[String](this, "selectionMode",
    "The selection mode. Supported options: numTopFeatures, percentile, fpr, fdr, fwe",
    ParamValidators.inArray(Array("numTopFeatures", "percentile", "fpr", "fdr", "fwe")))

  /** @group getParam */
  @Since("3.1.1")
  def getSelectionMode: String = $(selectionMode)

  /**
   * The upper bound of the features that selector will select.
   * @group param
   */
  @Since("3.1.1")
  final val selectionThreshold = new DoubleParam(this, "selectionThreshold",
    "The upper bound of the features that selector will select.")

  /** @group getParam */
  def getSelectionThreshold: Double = $(selectionThreshold)

  setDefault(selectionMode -> "numTopFeatures")
}

/**
 * Feature selector based on univariate statistical tests against labels. Currently, Spark
 * supports three Univariate Feature Selectors: chi-squared, ANOVA F-test and F-value.
 * User can choose Univariate Feature Selector by setting `featureType` and `labelType`,
 * and Spark will pick the score function based on the specified `featureType` and `labelType`.
 *
 * The following combination of `featureType` and `labelType` are supported:
 *  - `featureType` `categorical` and `labelType` `categorical`: Spark uses chi-squared,
 *    i.e. chi2 in sklearn.
 *  - `featureType` `continuous` and `labelType` `categorical`: Spark uses ANOVA F-test,
 *    i.e. f_classif in sklearn.
 *  - `featureType` `continuous` and `labelType` `continuous`: Spark uses F-value,
 *    i.e. f_regression in sklearn.
 *
 * The `UnivariateFeatureSelector` supports different selection modes: `numTopFeatures`,
 * `percentile`, `fpr`, `fdr`, `fwe`.
 *  - `numTopFeatures` chooses a fixed number of top features according to a hypothesis.
 *  - `percentile` is similar but chooses a fraction of all features instead of a fixed number.
 *  - `fpr` chooses all features whose p-value are below a threshold, thus controlling the false
 *    positive rate of selection.
 *  - `fdr` uses the <a href=
 *  "https://en.wikipedia.org/wiki/False_discovery_rate#Benjamini.E2.80.93Hochberg_procedure">
 *  Benjamini-Hochberg procedure</a>
 *    to choose all features whose false discovery rate is below a threshold.
 *  - `fwe` chooses all features whose p-values are below a threshold. The threshold is scaled by
 *    1/numFeatures, thus controlling the family-wise error rate of selection.
 *
 * By default, the selection mode is `numTopFeatures`.
 */
@Since("3.1.1")
final class UnivariateFeatureSelector @Since("3.1.1")(@Since("3.1.1") override val uid: String)
  extends Estimator[UnivariateFeatureSelectorModel] with UnivariateFeatureSelectorParams
    with DefaultParamsWritable {

  @Since("3.1.1")
  def this() = this(Identifiable.randomUID("UnivariateFeatureSelector"))

  /** @group setParam */
  @Since("3.1.1")
  def setSelectionMode(value: String): this.type = set(selectionMode, value)

  /** @group setParam */
  @Since("3.1.1")
  def setSelectionThreshold(value: Double): this.type = set(selectionThreshold, value)

  /** @group setParam */
  @Since("3.1.1")
  def setFeatureType(value: String): this.type = set(featureType, value)

  /** @group setParam */
  @Since("3.1.1")
  def setLabelType(value: String): this.type = set(labelType, value)

  /** @group setParam */
  @Since("3.1.1")
  def setFeaturesCol(value: String): this.type = set(featuresCol, value)

  /** @group setParam */
  @Since("3.1.1")
  def setOutputCol(value: String): this.type = set(outputCol, value)

  /** @group setParam */
  @Since("3.1.1")
  def setLabelCol(value: String): this.type = set(labelCol, value)

  @Since("3.1.1")
  override def fit(dataset: Dataset[_]): UnivariateFeatureSelectorModel = {
    transformSchema(dataset.schema, logging = true)
    val numFeatures = DatasetUtils.getNumFeatures(dataset, $(featuresCol))

    var threshold = Double.NaN
    if (isSet(selectionThreshold)) {
      threshold = $(selectionThreshold)
    } else {
      $(selectionMode) match {
        case "numTopFeatures" => threshold = 50
        case "percentile" => threshold = 0.1
        case "fpr" | "fdr" | "fwe" => threshold = 0.05
      }
    }

    val resultDF = ($(featureType), $(labelType)) match {
      case ("categorical", "categorical") =>
        ChiSquareTest.test(dataset.toDF, getFeaturesCol, getLabelCol, true)
      case ("continuous", "categorical") =>
        ANOVATest.test(dataset.toDF, getFeaturesCol, getLabelCol, true)
      case ("continuous", "continuous") =>
        FValueTest.test(dataset.toDF, getFeaturesCol, getLabelCol, true)
      case _ =>
        throw new IllegalArgumentException(s"Unsupported combination:" +
          s" featureType=${$(featureType)}, labelType=${$(labelType)}")
    }

    val indices = selectIndicesFromPValues(numFeatures, resultDF, $(selectionMode), threshold)
    copyValues(new UnivariateFeatureSelectorModel(uid, indices)
      .setParent(this))
  }

  private def getTopIndices(df: DataFrame, k: Int): Array[Int] = {
    val spark = SparkSession.builder().getOrCreate()
    import spark.implicits._
    df.sort("pValue", "featureIndex")
      .select("featureIndex")
      .limit(k)
      .as[Int]
      .collect()
  }

  private[feature] def selectIndicesFromPValues(
      numFeatures: Int,
      resultDF: DataFrame,
      selectionMode: String,
      selectionThreshold: Double): Array[Int] = {
    val spark = SparkSession.builder().getOrCreate()
    import spark.implicits._
    val indices = selectionMode match {
      case "numTopFeatures" =>
        getTopIndices(resultDF, selectionThreshold.toInt)
      case "percentile" =>
        getTopIndices(resultDF, (numFeatures * selectionThreshold).toInt)
      case "fpr" =>
        resultDF.select("featureIndex")
          .where(col("pValue") < selectionThreshold)
          .as[Int].collect()
      case "fdr" =>
        // This uses the Benjamini-Hochberg procedure.
        // https://en.wikipedia.org/wiki/False_discovery_rate#Benjamini.E2.80.93Hochberg_procedure
        val f = selectionThreshold / numFeatures
        val maxIndex = resultDF.sort("pValue", "featureIndex")
          .select("pValue")
          .as[Double].rdd
          .zipWithIndex
          .flatMap { case (pValue, index) =>
            if (pValue <= f * (index + 1)) {
              Iterator.single(index.toInt)
            } else Iterator.empty
          }.fold(-1)(math.max)
        if (maxIndex >= 0) {
          getTopIndices(resultDF, maxIndex + 1)
        } else Array.emptyIntArray
      case "fwe" =>
        resultDF.select("featureIndex")
          .where(col("pValue") < selectionThreshold / numFeatures)
          .as[Int].collect()
      case errorType =>
        throw new IllegalArgumentException(s"Unknown Selector Type: $errorType")
    }
    indices
  }

  @Since("3.1.1")
  override def transformSchema(schema: StructType): StructType = {
    if (isSet(selectionThreshold)) {
      val threshold = $(selectionThreshold)
      $(selectionMode) match {
        case "numTopFeatures" =>
          require(threshold >= 1 && threshold.toInt == threshold,
            s"selectionThreshold needs to be a positive Integer for selection mode " +
              s"numTopFeatures, but got $threshold")
        case "percentile" | "fpr" | "fdr" | "fwe" =>
          require(0 <= threshold && threshold <= 1,
            s"selectionThreshold needs to be in the range [0, 1] for selection mode " +
              s"${$(selectionMode)}, but got $threshold")
      }
    }
    require(isSet(featureType) && isSet(labelType), "featureType and labelType need to be set")
    SchemaUtils.checkColumnType(schema, $(featuresCol), new VectorUDT)
    SchemaUtils.checkNumericType(schema, $(labelCol))
    SchemaUtils.appendColumn(schema, $(outputCol), new VectorUDT)
  }

  @Since("3.1.1")
  override def copy(extra: ParamMap): UnivariateFeatureSelector = defaultCopy(extra)
}

@Since("3.1.1")
object UnivariateFeatureSelector extends DefaultParamsReadable[UnivariateFeatureSelector] {

  @Since("3.1.1")
  override def load(path: String): UnivariateFeatureSelector = super.load(path)
}

/**
 * Model fitted by [[UnivariateFeatureSelectorModel]].
 */
@Since("3.1.1")
class UnivariateFeatureSelectorModel private[ml](
    @Since("3.1.1") override val uid: String,
    @Since("3.1.1") val selectedFeatures: Array[Int])
  extends Model[UnivariateFeatureSelectorModel] with UnivariateFeatureSelectorParams
    with MLWritable {

  /** @group setParam */
  @Since("3.1.1")
  def setFeaturesCol(value: String): this.type = set(featuresCol, value)

  /** @group setParam */
  @Since("3.1.1")
  def setOutputCol(value: String): this.type = set(outputCol, value)

  protected def isNumericAttribute = true

  @Since("3.1.1")
  override def transform(dataset: Dataset[_]): DataFrame = {
    val outputSchema = transformSchema(dataset.schema, logging = true)

    UnivariateFeatureSelectorModel
      .transform(dataset, selectedFeatures.sorted, outputSchema, $(outputCol), $(featuresCol))
  }

  @Since("3.1.1")
  override def transformSchema(schema: StructType): StructType = {
    SchemaUtils.checkColumnType(schema, $(featuresCol), new VectorUDT)
    val newField =
      UnivariateFeatureSelectorModel
        .prepOutputField(schema, selectedFeatures, $(outputCol), $(featuresCol), isNumericAttribute)
    SchemaUtils.appendColumn(schema, newField)
  }

  @Since("3.1.1")
  override def copy(extra: ParamMap): UnivariateFeatureSelectorModel = {
    val copied = new UnivariateFeatureSelectorModel(uid, selectedFeatures)
      .setParent(parent)
    copyValues(copied, extra)
  }

  @Since("3.1.1")
  override def write: MLWriter =
    new UnivariateFeatureSelectorModel.UnivariateFeatureSelectorModelWriter(this)

  @Since("3.1.1")
  override def toString: String = {
    s"UnivariateFeatureSelectorModel: uid=$uid, numSelectedFeatures=${selectedFeatures.length}"
  }
}

@Since("3.1.1")
object UnivariateFeatureSelectorModel extends MLReadable[UnivariateFeatureSelectorModel] {

  @Since("3.1.1")
  override def read: MLReader[UnivariateFeatureSelectorModel] =
    new UnivariateFeatureSelectorModelReader

  @Since("3.1.1")
  override def load(path: String): UnivariateFeatureSelectorModel = super.load(path)

  private[UnivariateFeatureSelectorModel] class UnivariateFeatureSelectorModelWriter(
      instance: UnivariateFeatureSelectorModel) extends MLWriter {

    private case class Data(selectedFeatures: Seq[Int])

    override protected def saveImpl(path: String): Unit = {
      DefaultParamsWriter.saveMetadata(instance, path, sc)
      val data = Data(instance.selectedFeatures.toSeq)
      val dataPath = new Path(path, "data").toString
      sparkSession.createDataFrame(Seq(data)).repartition(1).write.parquet(dataPath)
    }
  }

  private class UnivariateFeatureSelectorModelReader
    extends MLReader[UnivariateFeatureSelectorModel] {

    /** Checked against metadata when loading model */
    private val className = classOf[UnivariateFeatureSelectorModel].getName

    override def load(path: String): UnivariateFeatureSelectorModel = {
      val metadata = DefaultParamsReader.loadMetadata(path, sc, className)
      val dataPath = new Path(path, "data").toString
      val data = sparkSession.read.parquet(dataPath)
        .select("selectedFeatures").head()
      val selectedFeatures = data.getAs[Seq[Int]](0).toArray
      val model = new UnivariateFeatureSelectorModel(metadata.uid, selectedFeatures)
      metadata.getAndSetParams(model)
      model
    }
  }

  private def transform(
      dataset: Dataset[_],
      selectedFeatures: Array[Int],
      outputSchema: StructType,
      outputCol: String,
      featuresCol: String): DataFrame = {
    val newSize = selectedFeatures.length
    val func = { vector: Vector =>
      vector match {
        case SparseVector(_, indices, values) =>
          val (newIndices, newValues) =
            compressSparse(indices, values, selectedFeatures)
          Vectors.sparse(newSize, newIndices, newValues)
        case DenseVector(values) =>
          Vectors.dense(selectedFeatures.map(values))
        case other =>
          throw new UnsupportedOperationException(
            s"Only sparse and dense vectors are supported but got ${other.getClass}.")
      }
    }

    val transformer = udf(func)
    dataset.withColumn(outputCol, transformer(col(featuresCol)),
      outputSchema(outputCol).metadata)
  }

  /**
   * Prepare the output column field, including per-feature metadata.
   */
  private def prepOutputField(
      schema: StructType,
      selectedFeatures: Array[Int],
      outputCol: String,
      featuresCol: String,
      isNumericAttribute: Boolean): StructField = {
    val selector = selectedFeatures.toSet
    val origAttrGroup = AttributeGroup.fromStructField(schema(featuresCol))
    val featureAttributes: Array[Attribute] = if (origAttrGroup.attributes.nonEmpty) {
      origAttrGroup.attributes.get.zipWithIndex.filter(x => selector.contains(x._2)).map(_._1)
    } else {
      if (isNumericAttribute) {
        Array.fill[Attribute](selector.size)(NumericAttribute.defaultAttr)
      } else {
        Array.fill[Attribute](selector.size)(NominalAttribute.defaultAttr)
      }
    }
    val newAttributeGroup = new AttributeGroup(outputCol, featureAttributes)
    newAttributeGroup.toStructField()
  }

  private def compressSparse(
      indices: Array[Int],
      values: Array[Double],
      selectedFeatures: Array[Int]): (Array[Int], Array[Double]) = {
    val newValues = new ArrayBuilder.ofDouble
    val newIndices = new ArrayBuilder.ofInt
    var i = 0
    var j = 0
    while (i < indices.length && j < selectedFeatures.length) {
      val indicesIdx = indices(i)
      val filterIndicesIdx = selectedFeatures(j)
      if (indicesIdx == filterIndicesIdx) {
        newIndices += j
        newValues += values(i)
        j += 1
        i += 1
      } else {
        if (indicesIdx > filterIndicesIdx) {
          j += 1
        } else {
          i += 1
        }
      }
    }
    // TODO: Sparse representation might be ineffective if (newSize ~= newValues.size)
    (newIndices.result(), newValues.result())
  }
}

相关信息

spark 源码目录

相关文章

spark Binarizer 源码

spark BucketedRandomProjectionLSH 源码

spark Bucketizer 源码

spark ChiSqSelector 源码

spark CountVectorizer 源码

spark DCT 源码

spark ElementwiseProduct 源码

spark FeatureHasher 源码

spark HashingTF 源码

spark IDF 源码