1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
|
/*
* 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.regression
import com.github.fommil.netlib.BLAS.{getInstance => blas}
import org.apache.spark.annotation.{Experimental, Since}
import org.apache.spark.internal.Logging
import org.apache.spark.ml.{PredictionModel, Predictor}
import org.apache.spark.ml.param.{Param, ParamMap}
import org.apache.spark.ml.tree.{GBTParams, TreeEnsembleModel, TreeRegressorParams}
import org.apache.spark.ml.tree.impl.GradientBoostedTrees
import org.apache.spark.ml.util.{Identifiable, MetadataUtils}
import org.apache.spark.mllib.linalg.Vector
import org.apache.spark.mllib.regression.LabeledPoint
import org.apache.spark.mllib.tree.configuration.{Algo => OldAlgo}
import org.apache.spark.mllib.tree.loss.{AbsoluteError => OldAbsoluteError, Loss => OldLoss,
SquaredError => OldSquaredError}
import org.apache.spark.mllib.tree.model.{GradientBoostedTreesModel => OldGBTModel}
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.functions._
/**
* :: Experimental ::
* [[http://en.wikipedia.org/wiki/Gradient_boosting Gradient-Boosted Trees (GBTs)]]
* learning algorithm for regression.
* It supports both continuous and categorical features.
*/
@Since("1.4.0")
@Experimental
final class GBTRegressor @Since("1.4.0") (@Since("1.4.0") override val uid: String)
extends Predictor[Vector, GBTRegressor, GBTRegressionModel]
with GBTParams with TreeRegressorParams with Logging {
@Since("1.4.0")
def this() = this(Identifiable.randomUID("gbtr"))
// Override parameter setters from parent trait for Java API compatibility.
// Parameters from TreeRegressorParams:
@Since("1.4.0")
override def setMaxDepth(value: Int): this.type = super.setMaxDepth(value)
@Since("1.4.0")
override def setMaxBins(value: Int): this.type = super.setMaxBins(value)
@Since("1.4.0")
override def setMinInstancesPerNode(value: Int): this.type =
super.setMinInstancesPerNode(value)
@Since("1.4.0")
override def setMinInfoGain(value: Double): this.type = super.setMinInfoGain(value)
@Since("1.4.0")
override def setMaxMemoryInMB(value: Int): this.type = super.setMaxMemoryInMB(value)
@Since("1.4.0")
override def setCacheNodeIds(value: Boolean): this.type = super.setCacheNodeIds(value)
@Since("1.4.0")
override def setCheckpointInterval(value: Int): this.type = super.setCheckpointInterval(value)
/**
* The impurity setting is ignored for GBT models.
* Individual trees are built using impurity "Variance."
*/
@Since("1.4.0")
override def setImpurity(value: String): this.type = {
logWarning("GBTRegressor.setImpurity should NOT be used")
this
}
// Parameters from TreeEnsembleParams:
@Since("1.4.0")
override def setSubsamplingRate(value: Double): this.type = super.setSubsamplingRate(value)
@Since("1.4.0")
override def setSeed(value: Long): this.type = super.setSeed(value)
// Parameters from GBTParams:
@Since("1.4.0")
override def setMaxIter(value: Int): this.type = super.setMaxIter(value)
@Since("1.4.0")
override def setStepSize(value: Double): this.type = super.setStepSize(value)
// Parameters for GBTRegressor:
/**
* Loss function which GBT tries to minimize. (case-insensitive)
* Supported: "squared" (L2) and "absolute" (L1)
* (default = squared)
* @group param
*/
@Since("1.4.0")
val lossType: Param[String] = new Param[String](this, "lossType", "Loss function which GBT" +
" tries to minimize (case-insensitive). Supported options:" +
s" ${GBTRegressor.supportedLossTypes.mkString(", ")}",
(value: String) => GBTRegressor.supportedLossTypes.contains(value.toLowerCase))
setDefault(lossType -> "squared")
/** @group setParam */
@Since("1.4.0")
def setLossType(value: String): this.type = set(lossType, value)
/** @group getParam */
@Since("1.4.0")
def getLossType: String = $(lossType).toLowerCase
/** (private[ml]) Convert new loss to old loss. */
override private[ml] def getOldLossType: OldLoss = {
getLossType match {
case "squared" => OldSquaredError
case "absolute" => OldAbsoluteError
case _ =>
// Should never happen because of check in setter method.
throw new RuntimeException(s"GBTRegressorParams was given bad loss type: $getLossType")
}
}
override protected def train(dataset: DataFrame): GBTRegressionModel = {
val categoricalFeatures: Map[Int, Int] =
MetadataUtils.getCategoricalFeatures(dataset.schema($(featuresCol)))
val oldDataset: RDD[LabeledPoint] = extractLabeledPoints(dataset)
val numFeatures = oldDataset.first().features.size
val boostingStrategy = super.getOldBoostingStrategy(categoricalFeatures, OldAlgo.Regression)
val (baseLearners, learnerWeights) = GradientBoostedTrees.run(oldDataset, boostingStrategy,
$(seed))
new GBTRegressionModel(uid, baseLearners, learnerWeights, numFeatures)
}
@Since("1.4.0")
override def copy(extra: ParamMap): GBTRegressor = defaultCopy(extra)
}
@Since("1.4.0")
@Experimental
object GBTRegressor {
// The losses below should be lowercase.
/** Accessor for supported loss settings: squared (L2), absolute (L1) */
@Since("1.4.0")
final val supportedLossTypes: Array[String] = Array("squared", "absolute").map(_.toLowerCase)
}
/**
* :: Experimental ::
*
* [[http://en.wikipedia.org/wiki/Gradient_boosting Gradient-Boosted Trees (GBTs)]]
* model for regression.
* It supports both continuous and categorical features.
* @param _trees Decision trees in the ensemble.
* @param _treeWeights Weights for the decision trees in the ensemble.
*/
@Since("1.4.0")
@Experimental
final class GBTRegressionModel private[ml](
override val uid: String,
private val _trees: Array[DecisionTreeRegressionModel],
private val _treeWeights: Array[Double],
override val numFeatures: Int)
extends PredictionModel[Vector, GBTRegressionModel]
with TreeEnsembleModel[DecisionTreeRegressionModel] with Serializable {
require(_trees.nonEmpty, "GBTRegressionModel requires at least 1 tree.")
require(_trees.length == _treeWeights.length, "GBTRegressionModel given trees, treeWeights of" +
s" non-matching lengths (${_trees.length}, ${_treeWeights.length}, respectively).")
/**
* Construct a GBTRegressionModel
* @param _trees Decision trees in the ensemble.
* @param _treeWeights Weights for the decision trees in the ensemble.
*/
@Since("1.4.0")
def this(uid: String, _trees: Array[DecisionTreeRegressionModel], _treeWeights: Array[Double]) =
this(uid, _trees, _treeWeights, -1)
@Since("1.4.0")
override def trees: Array[DecisionTreeRegressionModel] = _trees
@Since("1.4.0")
override def treeWeights: Array[Double] = _treeWeights
override protected def transformImpl(dataset: DataFrame): DataFrame = {
val bcastModel = dataset.sqlContext.sparkContext.broadcast(this)
val predictUDF = udf { (features: Any) =>
bcastModel.value.predict(features.asInstanceOf[Vector])
}
dataset.withColumn($(predictionCol), predictUDF(col($(featuresCol))))
}
override protected def predict(features: Vector): Double = {
// TODO: When we add a generic Boosting class, handle transform there? SPARK-7129
// Classifies by thresholding sum of weighted tree predictions
val treePredictions = _trees.map(_.rootNode.predictImpl(features).prediction)
blas.ddot(numTrees, treePredictions, 1, _treeWeights, 1)
}
/** Number of trees in ensemble */
val numTrees: Int = trees.length
@Since("1.4.0")
override def copy(extra: ParamMap): GBTRegressionModel = {
copyValues(new GBTRegressionModel(uid, _trees, _treeWeights, numFeatures),
extra).setParent(parent)
}
@Since("1.4.0")
override def toString: String = {
s"GBTRegressionModel (uid=$uid) with $numTrees trees"
}
/**
* Estimate of the importance of each feature.
*
* Each feature's importance is the average of its importance across all trees in the ensemble
* The importance vector is normalized to sum to 1. This method is suggested by Hastie et al.
* (Hastie, Tibshirani, Friedman. "The Elements of Statistical Learning, 2nd Edition." 2001.)
* and follows the implementation from scikit-learn.
*
* @see [[DecisionTreeRegressionModel.featureImportances]]
*/
@Since("2.0.0")
lazy val featureImportances: Vector = TreeEnsembleModel.featureImportances(trees, numFeatures)
/** (private[ml]) Convert to a model in the old API */
private[ml] def toOld: OldGBTModel = {
new OldGBTModel(OldAlgo.Regression, _trees.map(_.toOld), _treeWeights)
}
}
private[ml] object GBTRegressionModel {
/** (private[ml]) Convert a model from the old API */
def fromOld(
oldModel: OldGBTModel,
parent: GBTRegressor,
categoricalFeatures: Map[Int, Int],
numFeatures: Int = -1): GBTRegressionModel = {
require(oldModel.algo == OldAlgo.Regression, "Cannot convert GradientBoostedTreesModel" +
s" with algo=${oldModel.algo} (old API) to GBTRegressionModel (new API).")
val newTrees = oldModel.trees.map { tree =>
// parent for each tree is null since there is no good way to set this.
DecisionTreeRegressionModel.fromOld(tree, null, categoricalFeatures)
}
val uid = if (parent != null) parent.uid else Identifiable.randomUID("gbtr")
new GBTRegressionModel(uid, newTrees, oldModel.treeWeights, numFeatures)
}
}
|