算法小白的第一次尝试---LDA(线性判别分析)降维 【适用于任何维度】

-------------------------------------------------------------------------------------
笔者追求算法实现，不喜欢大篇幅叙述原理，有关LDA(线性判别分析)理论推荐查看该篇博客

https://www.cnblogs.com/pinard/p/6244265.html
-------------------------------------------------------------------------------------

import breeze.linalg.DenseMatrix
import org.apache.log4j.{Level, Logger}
import org.apache.spark.ml.feature.{LabeledPoint,VectorAssembler}
import org.apache.spark.ml.linalg.Vectors
import org.apache.spark.sql.types.{DoubleType, StructField, StructType}
import org.apache.spark.sql.{DataFrame, Row, SparkSession}
import scala.collection.mutable.ArrayBuffer

/** The method is Linear discriminant analysis  which can be used to
  * lower the dimension of linear dataset
  *  Data Source :http://archive.ics.uci.edu/ml/datasets/Wine
  * @author XiaoTangBao
  * @date 2019/4/24 10:32
  * @version 1.0
  */
object LDA {
  def main(args: Array[String]): Unit = {
    //屏蔽日志
    Logger.getLogger("org.apache.spark").setLevel(Level.ERROR)
    //spark初始化
    val spark = SparkSession.builder().master("local[4]").appName("LDA").getOrCreate()
    //获取数据源   http://archive.ics.uci.edu/ml/datasets/Wine
    val data = spark.sparkContext.textFile("G:\\mldata\\wine.data").map(line => line.split(","))
          .map(arr => arr.map(str => str.toDouble)).map(arr =>Row(arr(0),arr(1),arr(2),arr(3),arr(4),arr(5),
          arr(6),arr(7),arr(8),arr(9),arr(10),arr(11),arr(12),arr(13)))

    //设置featuresArr和schema,便于后期数据转化及生成dataFrame
    val featuresArr = Array("Alcohol","Malic acid","Ash","Alcalinity of ash","Magnesium",
      "Total phenols","Flavanoids","Nonflavanoid phenols","Proanthocyanins","Color intensity",
      "Hue","OD280/OD315 of diluted wines","Proline")
    val schema = StructType(List(StructField("label",DoubleType,true),StructField("Alcohol",DoubleType,true),StructField("Malic acid",DoubleType,true),
      StructField("Ash",DoubleType,true),StructField("Alcalinity of ash",DoubleType,true),StructField("Magnesium",DoubleType,true)
      ,StructField("Total phenols",DoubleType,true),StructField("Flavanoids",DoubleType,true),StructField("Nonflavanoid phenols",DoubleType,true)
      ,StructField("Proanthocyanins",DoubleType,true),StructField("Color intensity",DoubleType,true),StructField("Hue",DoubleType,true)
      ,StructField("OD280/OD315 of diluted wines",DoubleType,true),StructField("Proline",DoubleType,true)))
    val oridf = spark.createDataFrame(data,schema)

    //设置转化器
    val vectorAsb = new VectorAssembler().setInputCols(featuresArr).setOutputCol("features")
    //数据整理后传入run，启动LDA算法
    val newdf = vectorAsb.transform(oridf).select("label","features")
    val rpg = run(newdf,2)
    val arr = ArrayBuffer[(Double,Double)]()
    for(i<-0 until rpg.cols) arr.append((rpg(0,i),rpg(1,i)))
    arr.foreach(tp =>println(tp._1))
    println()
    arr.foreach(tp =>println(tp._2))
  }

  /**
    * Entrance of modeltraining
    * @param df trainData with only two columns such as label and features
    * @param nb the dimensions of the traindata after
    */
 def run(df:DataFrame,nb:Int)={
   val trainData = df.select("features").rdd.map(row => row.toString())
    .map(str => str.replace('[',' '))
    .map(str => str.replace(']',' '))
    .map(str => str.trim).map(str => str.split(','))
    .map(arr => arr.map(str => str.toDouble)).collect()

   val labels = df.select("label").rdd.map(row => row.toString())
     .map(str => str.replace('[',' '))
     .map(str => str.replace(']',' '))
     .map(str => str.trim).map(str => str.toDouble).collect()
   //特征列数
   val tz = trainData(0).length
   //生成新的带label的数据
   val labArr = ArrayBuffer[LabeledPoint]()
   for(i<-0 until trainData.length) labArr.append(LabeledPoint(labels(i),Vectors.dense(trainData(i))))
   //总样本组成的大型矩阵
   val allData = labArr.map(lab => lab.features).map(vec => vec.toArray).flatMap(x => x).toArray
   val big_Matrx =new DenseMatrix[Double](tz,trainData.length,allData)
   import breeze.linalg._
   //存放向量各维度的均值
   val big_mean = sum(big_Matrx,Axis._1).*= (1.0 / big_Matrx.cols)
   //总的类别
   val allLabel = labels.distinct
   //类内散度矩阵
   val Sw_Arr = ArrayBuffer[DenseMatrix[Double]]()
   //类间散度矩阵
   val Sb_Arr = ArrayBuffer[DenseMatrix[Double]]()
   for(i<-0 until allLabel.length){
     //该类别下的总记录数
     val record = labArr.filter(lab => lab.label == allLabel(i)).size
     val sk = labArr.filter(lab => lab.label == allLabel(i)).map(lab => lab.features)
       .map(vec => vec.toArray).flatMap(x => x).toArray
     var d1 = new DenseMatrix[Double](tz,record,sk)
     //存放向量各维度的均值
     val cols_mean = sum(d1,Axis._1).*= (1.0 / d1.cols)
     //样本去中心化
     for(i<-0 until d1.cols){
        d1(::,i) := d1(::,i) - cols_mean
     }
     //类内散度矩阵
     val sw = d1 * (d1.t)
      Sw_Arr.append(sw)
     //类间散度矩阵
     val zf = (cols_mean - big_mean).toDenseMatrix.t
     val sb = record.toDouble * zf * zf.t
     Sb_Arr.append(sb)
   }
   //总类内散度矩阵
   var total_Sw = DenseMatrix.zeros[Double](tz,tz)
   for(i<-0 until Sw_Arr.length) total_Sw = total_Sw + Sw_Arr(i)
   //总类间散度矩阵
   var total_Sb = DenseMatrix.zeros[Double](tz,tz)
   for(i<-0 until Sb_Arr.length) total_Sb = total_Sb + Sb_Arr(i)

   //计算类内散度和类间散度矩阵乘积
   val Sw_Sb = inv(total_Sw) * total_Sb
   //计算Sw_Sb矩阵特征值及特征向量
   val eigValues = eig(Sw_Sb).eigenvalues
   val eigVectors = eig(Sw_Sb).eigenvectors
   //测试结果表明，特征向量为单列向量，一列代表的才是一个特征向量，所以之前的理解是错的
   //选取最大的k个特征值对应的特征向量
   val label_eig = DenseMatrix.horzcat(eigVectors.t,eigValues.toDenseMatrix.t)
   var strArr = ArrayBuffer[String]()
   for(i<-0 until label_eig.rows) strArr.append(label_eig.t(::,i).toString)
   for(i<-0 until strArr.length){
     strArr(i) = strArr(i).replace("DenseVector(","").replace(')',' ').trim()
   }
   val da = ArrayBuffer[LabeledPoint]()
   for(str <- strArr){
     val arr = str.split(',').map(string => string.toDouble)
     val lab = arr.takeRight(1)(0)
     val value = arr.take(arr.length -1)
     val labPoint = LabeledPoint(lab,Vectors.dense(value))
     da.append(labPoint)
   }
  //rt表示最终选取的特征向量矩阵
   val result = da.sortBy(labPoint => labPoint.label).reverse.take(nb).map(lab => lab.features).map(vec => vec.toArray)
   var rt = DenseMatrix.zeros[Double](result.length,result(0).length)
   for(i<-0 until rt.rows){
     for(j<-0 until rt.cols){
       rt(i,j) = result(i)(j)
     }
   }

   //降维后的数据集
  val lastData = rt * big_Matrx
   lastData
 }
}

根据实验结果数据绘制图像如下图所示：

该结果与Python 直接调取LDA方法结果相差较大：

这是由于Spark和python求取的取特征向量不同导致，因为矩阵特征向量本身就非唯一，
同一特征值对应的特征向量有无数个，将Spark求取的第二个特征向量乘以-1后，结果如下：

此时可以发现，该结果与Python调包结果几乎一致，横纵座标的不同依旧是由于特征向量的不完全一致导致的。因为作者仅仅改变了向量的方向，并没有对向量进行缩放。