scala實現kmeans算法

算法的概念不做過都解釋，google一下一大把。直接貼上代碼，有比較詳細的注釋了。

主程序：

import scala.io.Source
import scala.util.Random

/**
 * @author vincent
 *
 */
object LocalKMeans {
    def main(args: Array[String]) {
        val fileName = "/home/vincent/kmeans_data.txt"
        val knumbers = 3
        val rand = new Random()

        //  讀取文本數據
        val lines = Source.fromFile(fileName).getLines.toArray
        val points = lines.map(line => {
            val parts = line.split("\t").map(_.toDouble)
            new Point(parts(0), parts(1))
        }).toArray
        
        //  隨機初始化k個質心
        val centroids = new Array[Point](knumbers)
        for (i <- 0 until knumbers) {
            centroids(i) = points(new Random().nextInt(points.length))
        }
        val startTime = System.currentTimeMillis()
        println("initialize centroids:\n" + centroids.mkString("\n") + "\n")
        println("test points: \n" + points.mkString("\n") + "\n")

        val resultCentroids = kmeans(points, centroids, 0.001)
        
        val endTime = System.currentTimeMillis()
        val runTime = endTime - startTime
        println("run Time: " + runTime + "\nFinal centroids: \n" + resultCentroids.mkString("\n"))
    }
    
    //  算法的核心函數
    def kmeans(points: Seq[Point], centroids: Seq[Point], epsilon: Double): Seq[Point] = {
        //  最近質心為key值，將數據集分簇
        val clusters = points.groupBy(closestCentroid(centroids, _))
        println("clusters: \n" + clusters.mkString("\n") + "\n")
        //  分別計算簇中數據集的平均數，得到每個簇的新質心
        val newCentroids = centroids.map(oldCentroid => {
            clusters.get(oldCentroid) match {
                case Some(pointsInCluster) => pointsInCluster.reduceLeft(_ + _) / pointsInCluster.length
                case None => oldCentroid
            }
        })
        //  計算新質心相對與舊質心的偏移量
        val movement = (centroids zip newCentroids).map({ case (a, b) => a distance b })
        println("Centroids changed by\n" + movement.map(d => "%3f".format(d)).mkString("(", ", ", ")")
            + "\nto\n" + newCentroids.mkString(", ") + "\n")
        //  根據偏移值大小決定是否繼續迭代，epsilon為最小偏移值
        if (movement.exists(_ > epsilon))
            kmeans(points, newCentroids, epsilon)
        else
            return newCentroids
    }

    //  計算最近質心
    def closestCentroid(centroids: Seq[Point], point: Point) = {
        centroids.reduceLeft((a, b) => if ((point distance a) < (point distance b)) a else b)
    }
}

 

自定義Point類：

/**
 * @author vincent
 *
 */
object Point {
    def random() = {
        new Point(math.random * 50, math.random * 50)
    }
}

case class Point(val x: Double, val y: Double) {
    def +(that: Point) = new Point(this.x + that.x, this.y + that.y)
    def -(that: Point) = new Point(this.x - that.x, this.y - that.y)
    def /(d: Double) = new Point(this.x / d, this.y / d)
    def pointLength = math.sqrt(x * x + y * y)
    def distance(that: Point) = (this - that).pointLength
    override def toString = format("(%.3f, %.3f)", x, y)
}

測試數據集：

12.044996    36.412378
31.881257    33.677009
41.703139    46.170517
43.244406    6.991669
19.319000    27.926669
3.556824    40.935215
29.328655    33.303675
43.702858    22.305344
28.978940    28.905725
10.426760    40.311507