K近邻(kNN,k-NearestNeighbor)分类算法基本思想是:如果一个样本在特征空间中的k个最相似,也就是特征空间中k个最邻近的样本大多数属于某一个类别,则该样本也属于这个类别。类似与古话:近朱者赤,近墨者黑,背后自然也蕴藏着物以类聚,人以群分的思想!
算法步骤:
1.对数据进行归一化处理
2.求每个测试样本基于训练样本的k个最近临样本
3.k个最近临样本所属类别中最大的一个即位测试样本的类别。
优点:
1.容易理解,易于分类
2.适合多类别的分类问题
缺点:
1.每个测试样本需要与所有训练样本进行求距离,计算量大
2.当各类样本不平衡时,测试结果可能会趋向与样本数量多的那一类。
k值的选择:
k值过小,得到的近临数太少,使得分类精度低,同时放大了噪声的干扰;k值过大,当各类样本不平衡时,测试结果可能会趋向与样本数量多的那一类。k值的选择一般小于训练样本的平方根。
实例:
《机器学习实战》一书中手写数字识别例子:
训练样本: 32*32的0,1文本,每个文本代表一个手写数字,文本名中包含该文本所属的数字类别。点击下载
测试样本:同训练样本
代码:
from numpy import *
import operator
import os
def ReadData(trainDir, testDir):
trainFileList = os.listdir(trainDir)
testFileList = os.listdir(testDir)
numSamples = len(trainFileList)
trainX = zeros((numSamples, 1024))
trainY = []
for i in xrange(numSamples):
fileName = trainFileList[i]
trainX[i, :] = ReadImgData(trainDir + fileName)
label = int(fileName.split('_')[0])
trainY.append(label)
numSamples = len(testFileList)
testX = zeros((numSamples, 1024))
testY = []
for i in xrange(numSamples):
fileName = testFileList[i]
testX[i, :] = ReadImgData(testDir + fileName)
label = int(fileName.split('_')[0])
testY.append(label)
return trainX, trainY, testX, testY
#ReadImgData读取每个文本内容
def ReadImgData(fileName):
row = 32
col = 32
fileX = zeros((1, row*col))
fileFp = open(fileName)
for i in xrange(row):
lineTemp = fileFp.readline()
for j in xrange(col):
fileX[0, i*row + j] = int(lineTemp[j])
return fileX
def knn(testX, trainX, trainY, k):
numSamples = trainX.shape[0]
diff = tile(testX, (numSamples, 1)) - trainX
squareDiff = diff ** 2
squareDist = sum(squareDiff, axis = 1)
dist = squareDist ** 0.5
sortedDist = argsort(dist)
classCount = {}
for i in xrange(k):
voteLabel = trainY[sortedDist[i]]
classCount[voteLabel] = classCount.get(voteLabel, 0) + 1
maxCount = 0
for key, value in classCount.items():
if value > maxCount:
maxCount = value
maxIndex = key
return maxIndex
print "******start******"
trainDir = './trainingDigits/'
testDir = './testDigits/'
trainX, trainY, testX, testY = ReadData(trainDir, testDir)
print "******Data End******"
sumSamples = testX.shape[0]
right = 0
for i in xrange(sumSamples):
label = knn(testX[i], trainX, trainY, 3)
#print "label = %d" % label
if label == testY[i]:
right += 1
print "*****Test End******"
print 'right = %d' % right
rate = float(right) / sumSamples
print 'rate = %f' % rate
代码中ReadData读取样本目录中的数据,ReadImgData读取单个文本的数据,knn实现测试算法。