Spark -- Spark的smote算法，非平衡数据处理示例

介绍

  相关的理论，和代码可以参考：https://www.cnblogs.com/little-horse/p/11241168.html，这个参考代码我看了感觉有点偏老和偏复杂，于是找了另外一个版本。
  spark的issues:https://issues.apache.org/jira/browse/SPARK-18441，其中有一个示例：https://gist.github.com/hhbyyh/346467373014943a7f20df208caeb19b，该网址需要翻墙，如果需要可以往下看。

代码

  先说明下k和n的意义，k是指欧氏距离最近的k近邻，但只随机取其中一个近邻，以此为基准，在这条线上选取n个新样本。
  在引入下面代码后会报错，org.apache.spark.ml.linalg.BLAS是private，不能被直接访问，我这里直接新建了org.apache.spark.ml.linalg文件夹，新建BLAS类，全部复制过来去掉private就可以了。

github参考代码

spark issues的解答中的示例代码如下：

package org.apache.spark.ml.feature

import org.apache.spark.ml.linalg.BLAS.axpy
import org.apache.spark.ml.linalg._
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.SparkSession

import scala.util.Random

/**
  * Created by yuhao on 12/1/16.
  */
object SmoteSampler {

  def generateSamples(data: RDD[(Long, Vector)], k: Int, N: Int): RDD[Vector] = {

    val knei = data.cartesian(data).map { case ((id1, vec1), (id2, vec2)) =>
      (id1, vec1, vec2)
    }.groupBy(_._1)
      .map { case (id, iter) =>
        val arr = iter.toArray
        (arr(0)._2, arr.sortBy(t => Vectors.sqdist(t._2, t._3)).take(k + 1).tail.map(_._3))
      }

    knei.foreach(t => println(t._1 + "\t" + t._2.mkString(", ")))

    knei.flatMap { case (vec, neighbours) =>
      (1 to N).map { i =>
        val rn = neighbours(Random.nextInt(k))
        val diff = rn.copy
        axpy(-1.0, vec, diff)
        val newVec = vec.copy
        axpy(Random.nextDouble(), diff, newVec)
        newVec
      }.iterator
    }
  }

}

// put it in another file.
package org.apache.spark.ml.feature

import org.apache.spark.ml.linalg.BLAS.axpy
import org.apache.spark.ml.linalg._
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.SparkSession

import scala.util.Random

object SmoteTest {

  def main(args: Array[String]): Unit = {
    val spark = SparkSession
      .builder
      .master("local[2]")
      .appName("smote example")
      .getOrCreate()

    // $example on$
    val df = spark.createDataFrame(Seq(
      (0L, Vectors.dense(1, 2)),
      (1L, Vectors.dense(3, 4)),
      (2L, Vectors.dense(5, 6))
    )).toDF("id", "features")
    
    val k = 2
    val N = 3
    val data = df.rdd.map(r => (r.getLong(0), r.getAs[Vector](1)))
    val newSamples = SmoteSampler.generateSamples(data, k, N)
    newSamples.collect().foreach(println)
    spark.stop()
  }
}

  但是使用过后发现几个问题如下：

1. 代码中使用了rdd的笛卡尔积，数据稍微多一点点，笛卡尔积都是一个很大的值，我本地以5000样本测试，笛卡尔积就达到5000*5000=25000000，本地已经跑不出来结果了。
2. 上面的测试数据每个样本只给出了一个样本值，如果给出多个，例如给出多个label是0的，则上面的groupby，sort后取take和tail都会有问题。

修改后代码

  根据上面的问题，我首先对笛卡尔积那里做了修改，不再生成笛卡尔积了，而是在groupby之后对map中每个向量生成笛卡尔积，这样因为是分区操作而且是真正的操作数据，非常之快。并且对于同样label的多样本的排序和take都是没问题的。

def main(args: Array[String]): Unit = {
    val spark = SparkSession
      .builder
      .master("local[2]")
      .appName("smote example")
      .getOrCreate()

    // $example on$
    val df = spark.createDataFrame(Seq(
      (0L, Vectors.dense(1, 2)),
      (0L, Vectors.dense(3, 4)),
      (0L, Vectors.dense(11, 12)),
      (1L, Vectors.dense(5, 6)),
      (1L, Vectors.dense(7, 8)),
      (1L, Vectors.dense(9, 10))
    )).toDF("id", "features")

    val k = 2
    val N = 3
    val data = df.rdd.map(r => (r.getLong(0), r.getAs[Vector](1)))
    val newSamples = SmoteSampler.generateSamples1(data, k, N)
    newSamples.collect().foreach(println)
    spark.stop()
  } 


  def generateSamples1(data: RDD[(Long, Vector)], k: Int, N: Int): RDD[Vector] = {
    val groupedRDD = data.groupBy(_._1)
    val vecAndNeis: RDD[(Vector, Array[Vector])] = groupedRDD.flatMap { case (id, iter) =>
      val vecArr = iter.toArray.map(_._2)
      //对每个vector产生笛卡尔积
      val cartesianArr: Array[(Vector, Vector)] = vecArr.flatMap(vec1 => {
        vecArr.map(vec2 => (vec1, vec2))
      }).filter(tuple => tuple._1 != tuple._2)
      cartesianArr.groupBy(_._1).map { case (vec, vecArr) => {
        (vec, vecArr.sortBy(x => Vectors.sqdist(x._1, x._2)).take(k).map(_._2))
      }
      }
    }

    //1.从这k个近邻中随机挑选一个样本，以该随机样本为基准生成N个新样本
    val vecRDD = vecAndNeis.flatMap { case (vec, neighbours) =>
      (1 to N).map { i =>
        val rn = neighbours(Random.nextInt(k))
        val diff = rn.copy
        axpy(-1.0, vec, diff)
        val newVec = vec.copy
        axpy(Random.nextDouble(), diff, newVec)
        newVec
      }.iterator
    }
    vecRDD
  }

  可以看到产生结果是很快的，即使换上稍微大些数据也是很快的，而且同一个label多个样本也是没问题的，如果将第二个代码的测试数据拿到第一个测试，不是报错就是取的近邻有问题。

结果

根据我司需求的一个完整代码

  需求说明：只对指定的label值作为需要平衡的样本，对这些的每个样本选取k个近邻，取其中随机一个近邻为基准，生成n个新样本，最后合并到原数据中，并标识哪些是新增的，哪些是原有数据。

测试数据

http://archive.ics.uci.edu/ml/machine-learning-databases/00275/Bike-Sharing-Dataset.zip

hour.csv和day.csv都有如下属性，除了hour.csv文件中没有hr属性以外

• instant: 记录ID
• dteday : 时间日期
• season : 季节 (1:春季, 2:夏季, 3:秋季, 4:冬季)
• yr : 年份 (0: 2011, 1:2012)
• mnth : 月份 ( 1 to 12)
• hr : 当天时刻 (0 to 23)
• holiday : 当天是否是节假日(extracted from http://dchr.dc.gov/page/holiday-schedule)
• weekday : 周几
• workingday : 工作日 is 1, 其他 is 0.

• weathersit : 天气

• 1: Clear, Few clouds, Partly cloudy, Partly cloudy
• 2: Mist + Cloudy, Mist + Broken clouds, Mist + Few clouds, Mist
• 3: Light Snow, Light Rain + Thunderstorm + Scattered clouds, Light Rain + Scattered clouds
• 4: Heavy Rain + Ice Pallets + Thunderstorm + Mist, Snow + Fog
• temp : 气温 Normalized temperature in Celsius. The values are divided to 41 (max)
• atemp: 体感温度 Normalized feeling temperature in Celsius. The values are divided to 50 (max)
• hum: 湿度 Normalized humidity. The values are divided to 100 (max)
• windspeed: 风速Normalized wind speed. The values are divided to 67 (max)
• casual: 临时用户数count of casual users
• registered: 注册用户数count of registered users
• cnt: 目标变量，每小时的自行车的租用量，包括临时用户和注册用户count of total rental bikes including both casual and registered

完整代码

package com.bigblue.ml

import java.text.SimpleDateFormat
import java.util.Date

import org.apache.spark.ml.feature.VectorAssembler
import org.apache.spark.ml.linalg.BLAS.axpy
import org.apache.spark.ml.linalg._
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.functions._
import org.apache.spark.sql.{DataFrame, Row, SparkSession}

import scala.util.Random

/**
 * Created By TheBigBlue on 2020/3/23
 * Description : 
 */
object ImbalancedDataProcess {

  def main(args: Array[String]): Unit = {
    val spark: SparkSession = SparkSession.builder().appName("test-lightgbm").master("local[4]").getOrCreate()
    spark.sparkContext.setLogLevel("WARN")
    val originalData: DataFrame = spark.read.option("header", "true") //第一行作为Schema
      .option("inferSchema", "true") //推测schema类型
      //      .csv("/home/hdfs/hour.csv")
      .csv("file:///F:\\Cache\\Program\\TestData\\lightgbm/hour.csv")

    val kNei = 5
    val nNei = 10
    //少数样本值
    val minSample = 0
    //标签列
    val labelCol = "workingday"
    // 连续列
    val vecCols: Array[String] = Array("temp", "atemp", "hum", "casual", "cnt")

    import spark.implicits._
    //原始数据只保留label和features列，追加一列sign标识为老数据
    val inputDF = originalData.select(labelCol, vecCols: _*).withColumn("sign", lit("O"))
    //需要对最小样本值的数据处理
    val filteredDF = inputDF.filter($"$labelCol" === minSample)
    //合并为label和向量列
    val labelAndVecDF = new VectorAssembler().setInputCols(vecCols).setOutputCol("features")
      .transform(filteredDF).select(labelCol, "features")
    //转为rdd
    val inputRDD = labelAndVecDF.rdd.map(row => (row.get(0).toString.toLong, row.getAs[Vector](1)))
    println(inputRDD.count() + "\t" + new SimpleDateFormat("HH:mm:ss").format(new Date()))

    //smote算法
    val vecRDD: RDD[Vector] = SmoteSampler.generateSamples1(inputRDD, kNei, nNei)
    println(vecRDD.count() + "\t" + new SimpleDateFormat("HH:mm:ss").format(new Date()))

    //以下是公司要求的和之前数据合并
    //生成dataframe，将向量列展开，追加一列sign标识为新数据
    val vecDF: DataFrame = vecRDD.map(vec => (0, vec.toArray)).toDF(labelCol, "features")
    val newCols = (0 until vecCols.size).map(i => $"features".getItem(i).alias(vecCols(i)))
    //根据需求，新数据应该为样本量*n，当前测试数据label为0的样本量为5514，则会新增5514*10=55140
    val newDF = vecDF.select(($"$labelCol" +: newCols): _*).withColumn("sign", lit("N"))

    //和原数据合并
    val finalDF = inputDF.union(newDF)
    finalDF.show
    
    import scala.collection.JavaConversions._
    //查看原数据
    val aggSeq: Seq[Row] = originalData.groupBy(labelCol).agg(count(labelCol).as("labelCount"))
      .collectAsList().toSeq
    println(aggSeq)

    //查看平衡后数据，根据需求，则最终合并后，label为0的样本为55140+5514=60654
    val aggSeq1: Seq[Row] = finalDF.groupBy(labelCol).agg(count(labelCol).as("labelCount"))
      .collectAsList().toSeq
    println(aggSeq1)
  }

  //smote
  def generateSamples1(data: RDD[(Long, Vector)], k: Int, N: Int): RDD[Vector] = {
    val groupedRDD = data.groupBy(_._1)
    val vecAndNeis: RDD[(Vector, Array[Vector])] = groupedRDD.flatMap { case (id, iter) =>
      val vecArr = iter.toArray.map(_._2)
      //对每个vector产生笛卡尔积
      val cartesianArr: Array[(Vector, Vector)] = vecArr.flatMap(vec1 => {
        vecArr.map(vec2 => (vec1, vec2))
      }).filter(tuple => tuple._1 != tuple._2)
      cartesianArr.groupBy(_._1).map { case (vec, vecArr) => {
        (vec, vecArr.sortBy(x => Vectors.sqdist(x._1, x._2)).take(k).map(_._2))
      }
      }
    }

    //1.从这k个近邻中随机挑选一个样本，以该随机样本为基准生成N个新样本
    val vecRDD = vecAndNeis.flatMap { case (vec, neighbours) =>
      (1 to N).map { i =>
        val rn = neighbours(Random.nextInt(k))
        val diff = rn.copy
        axpy(-1.0, vec, diff)
        val newVec = vec.copy
        axpy(Random.nextDouble(), diff, newVec)
        newVec
      }.iterator
    }
    vecRDD
  }
}

结果

  可以看出平衡后，追加了一列标识，label为0的样本也增加了10倍，并且运行很快，在样本为5500左右，本地运行不到一分钟就出结果了。