再xgboost的源码中有xgboost的SparkWithDataFrame的实现，如下：https://github.com/dmlc/xgboost/tree/master/jvm-packages。但是由于各种各样的原因吧，这些代码在我的IDE里面编译不过，因此又写了如下代码以供以后查阅使用。
package xgboost
import ml.dmlc.xgboost4j.scala.spark.{XGBoost, XGBoostModel}
import org.apache.spark.mllib.evaluation.BinaryClassificationMetrics
import org.apache.spark.sql.{Row, DataFrame, SparkSession}
object App{
def main(args: Array[String]): Unit ={
val trainPath: String = "xxx/train.txt"
val testPath: String = "xxx/test.txt"
val binaryModelPath: String = "xxx/model.binary"
val textModelPath: String = "xxx/model.txt"
val spark = SparkSession
.builder()
.master("yarn")
.getOrCreate()
// define xgboost parameters
val maxDepth = 3
val numRound = 4
val nworker = 1
val paramMap = List(
"eta" -> 0.1,
"max_depth" -> maxDepth,
"objective" -> "binary:logistic").toMap
//read libsvm file
var dfTrain = spark.read.format("libsvm").load(trainPath).toDF("labelCol", "featureCol")
var dfTest = spark.read.format("libsvm").load(testPath).toDF("labelCol", "featureCol")
dfTrain.show(true)
printf("begin...")
val model:XGBoostModel = XGBoost.trainWithDataFrame(dfTrain, paramMap, numRound, nworker,
useExternalMemory = true,
featureCol = "featureCol", labelCol = "labelCol",
missing = 0.0f)
//predict the test set
val predict:DataFrame = model.transform(dfTest)
val scoreAndLabels = predict.select(model.getPredictionCol, model.getLabelCol)
.rdd
.map{case Row(score:Double, label:Double) => (score, label)}
//get the auc
val metric = new BinaryClassificationMetrics(scoreAndLabels)
val auc = metric.areaUnderROC()
println("auc:" + auc)
//save model
this.saveBinaryModel(model, spark, binaryModelPath)
this.saveTextModel(model, spark, textModelPath, numRound, maxDepth)
}
def saveBinaryModel(model:XGBoostModel, spark: SparkSession, path: String): Unit = {
model.saveModelAsHadoopFile(path)(spark.sparkContext)
}
def saveTextModel(model:XGBoostModel, spark: SparkSession, path: String, numRound: Int, maxDepth: Int): Unit = {
val dumpModel = model
.booster
.getModelDump()
.toList
.zipWithIndex
.map(x => s"booster:[${x._2}]\n${x._1}")
val header = s"numRound: $numRound, maxDepth: $maxDepth"
print(dumpModel)
import spark.implicits._
val text: List[String] = header +: dumpModel
text.toDF
.coalesce(1)
.write
.mode("overwrite")
.text(path)
}
}
其中：
1.训练集和测试集都是libsvm格式，如下所示：
1 3:1 10:1 11:1 21:1 30:1 34:1 36:1 40:1 41:1 53:1 58:1 65:1 69:1 77:1 86:1 88:1 92:1 95:1 102:1 105:1 117:1 124:1
0 3:1 10:1 20:1 21:1 23:1 34:1 36:1 39:1 41:1 53:1 56:1 65:1 69:1 77:1 86:1 88:1 92:1 95:1 102:1 106:1 116:1 120:1
2.最终生成的模型如下所示：
numRound: 4, maxDepth: 3 booster:[0] 0:[f29<2] yes=1,no=2,missing=2 1:leaf=0.152941 2:leaf=-0.191209 booster:[1] 0:[f29<2] yes=1,no=2,missing=2 1:leaf=0.141901 2:leaf=-0.174499 booster:[2] 0:[f29<2] yes=1,no=2,missing=2 1:leaf=0.132731 2:leaf=-0.161685 booster:[3] 0:[f29<2] yes=1,no=2,missing=2 1:leaf=0.124972 2:leaf=-0.15155
相关解释：”numRound: 4, maxDepth: 3”表示生成树的个数为4，树的最大深度为3；booster[n]表示第n棵树；以下保存树的结构，0号节点为根节点，每个节点有两个子节点，节点序号按层序技术，即1号和2号节点为根节点0号节点的子节点，相同层的节点有相同缩进，且比父节点多一级缩进。
在节点行，首先声明节点序号，中括号里写明该节点采用第几个特征（如f29即为训练数据的第29个特征），同时表明特征值划分条件，“[f29<2] yes=1,no=2,missing=2”：表示f29号特征大于2时该样本划分到1号叶子节点，f29>=2时划分到2号叶子节点，当没有该特征（None）划分到2号叶子节点。
3.预测的结果如下：
|labelCol|featureCol |probabilities |prediction| |1.0 |(126,[2,9,10,20,29,33,35,39,40,52,57,64,68,76,85,87,91,94,101,104,116,123],[1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0])|[0.3652743101119995,0.6347256898880005]|1.0 | |0.0 |(126,[2,9,19,20,22,33,35,38,40,52,55,64,68,76,85,87,91,94,101,105,115,119],[1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0])|[0.6635029911994934,0.3364970088005066]|0.0 |
