多分类实现方式介绍和在 Spark 上实现多分类逻辑回归

01

—

背景

03

—

OneVsRest介绍

OneVsRest将一个给定的二分类算法有效地扩展到多分类问题应用中，也叫做”One-vs-All”算法。OneVsRest是一个Estimator（评估器）。它采用一个基础的Classifier然后对于k个类别分别创建二分类问题。类别i的二分类分类器用来预测类别为i还是不为i，即将i类和其他类别区分开来。最后，通过依次对k个二分类分类器进行评估，取置信最高的分类器的标签作为i类别的标签。

对应多 “分类实现方式” 中的一对多

04

—

Spark中的多分类实现

基于ml包中的LogisticRegression实现

import org.apache.spark.sql.SparkSession
import org.apache.spark.ml.classification.{LogisticRegression, OneVsRest}
import org.apache.spark.ml.evaluation.MulticlassClassificationEvaluator

object MultiClassLR {
   def main(args: Array[String]): Unit = {
      val input_data = "data/sample_multiclass_classification_data.txt" // args(0)

      val spark = SparkSession.builder.master("local[5]").appName("MulticlassLRWithElasticNetExample")
         .getOrCreate()

         runBaseLR(spark,input_data)
      runBaseOneVsRest(spark,input_data)
      spark.stop()
   }

   def runBaseLR(spark: SparkSession, input_data: String): Unit = {
      // 加载训练数据集
      val split = spark.read.format("libsvm").load(input_data).randomSplit(Array(1,1))
      val train_data = split(0)
      val test_data = split(1)
      // 创建模型
      val lr = new LogisticRegression().setMaxIter(20).setRegParam(0.3).setElasticNetParam(0.8)
      // 训练模型
      val model = lr.fit(train_data)
      // 系数矩阵、截距向量
      println(s"coefficientMatrix is: \n  ${model.coefficientMatrix}")
      println(s"interceptVector is: \n  ${model.interceptVector}")

      // 测试集计算
      val predictions = model.transform(test_data)
      val test_count = test_data.count().toInt
      predictions.take(test_count).foreach(println)
      val evaluator = new MulticlassClassificationEvaluator()//.setLabelCol("label").setPredictionCol("prediction")
      val accuracy =evaluator.setMetricName("accuracy").evaluate(predictions);
      val weightedPrecision=evaluator.setMetricName("weightedPrecision").evaluate(predictions);
      val weightedRecall=evaluator.setMetricName("weightedRecall").evaluate(predictions);
      val f1=evaluator.setMetricName("f1").evaluate(predictions);

      println(s"accuracy is $accuracy")
      println(s"weightedPrecision is $weightedPrecision")
      println(s"weightedRecall is $weightedRecall")
      println(s"f1 is $f1")
   }
   def runBaseOneVsRest(spark: SparkSession, input_data: String): Unit = {
       // 加载训练数据集
      val split = spark.read.format("libsvm").load(input_data).randomSplit(Array(1,1))
      val train_data = split(0)
      val test_data = split(1)
      // 创建模型
      val lr = new LogisticRegression().setMaxIter(10).setRegParam(0.3).setElasticNetParam(0.8)
      // 训练模型
      val model = new OneVsRest().setClassifier(lr).fit(train_data)
         
      // 测试集计算
      val predictions = model.transform(test_data)
      val test_count = test_data.count().toInt
      predictions.take(test_count).foreach(println)
      val evaluator = new MulticlassClassificationEvaluator()//.setLabelCol("label").setPredictionCol("prediction")
      val accuracy =evaluator.setMetricName("accuracy").evaluate(predictions);
      val weightedPrecision=evaluator.setMetricName("weightedPrecision").evaluate(predictions);
      val weightedRecall=evaluator.setMetricName("weightedRecall").evaluate(predictions);
      val f1=evaluator.setMetricName("f1").evaluate(predictions);

      println(s"accuracy is $accuracy")
      println(s"weightedPrecision is $weightedPrecision")
      println(s"weightedRecall is $weightedRecall")
      println(s"f1 is $f1")
   }
}

运行输出信息（runBaseLR）

[0.0,(4,[0,1,2,3],[-0.666667,-0.583333,0.186441,0.333333]),[0.142419333934195,-0.3772619583435227,0.06140018515891296],[0.3973163522554911,0.23628803028122905,0.3663956174632798],0.0]
[0.0,(4,[0,1,2,3],[-0.277778,-0.333333,0.322034,0.583333]),[0.21570018327413792,-0.5776524462730686,0.06140018515891296],[0.433024032357093,0.19586792869116354,0.3711080389517435],0.0]
[0.0,(4,[0,1,2,3],[-0.222222,-0.583333,0.355932,0.583333]),[0.21570018327413792,-0.6014274236583748,0.06140018515891296],[0.43502594980545906,0.19215033820182928,0.3728237119927117],0.0]
[0.0,(4,[0,1,2,3],[-0.166667,-0.416667,0.38983,0.5]),[0.19127333120195608,-0.5901059157567916,0.06140018515891296],[0.42808566999671693,0.19596657170058288,0.37594775830270016],0.0]
...
accuracy is 0.8615384615384616
weightedPrecision is 0.9017369727047146
weightedRecall is 0.8615384615384616
f1 is 0.8554924320962056
...

运行输出信息（runBaseOneVsRest）

[0.0,(4,[0,1,2,3],[-0.666667,-0.583333,0.186441,0.333333]),0.0]
[0.0,(4,[0,1,2,3],[-0.222222,-0.583333,0.355932,0.583333]),0.0]
[0.0,(4,[0,1,2,3],[-0.111111,-0.166667,0.38983,0.416667]),0.0]
...
accuracy is 0.8051948051948052
weightedPrecision is 0.8539693389317449
weightedRecall is 0.8051948051948052
f1 is 0.7797931797931799
...