基于Python实现的ID3决策树功能示例_Python

本文实例讲述了基于Python实现的ID3 决策树功能。分享给大家供大家参考，具体如下：

ID3算法是决策树的一种，它是基于奥卡姆剃刀原理的，即用尽量用较少的东西做更多的事。ID3算法，即Iterative Dichotomiser 3，迭代二叉树3代，是Ross Quinlan发明的一种决策树算法，这个算法的基础就是上面提到的奥卡姆剃刀原理，越是小型的决策树越优于大的决策树，尽管如此，也不总是生成最小的树型结构，而是一个启发式算法。

如下示例是一个判断海洋生物数据是否是鱼类而构建的基于ID3思想的决策树

									# coding=utf-8

									import operator

									from math import log

									import time

									def createDataSet():

									  dataSet = [[1, 1, 'yes'],

									        [1, 1, 'yes'],

									        [1, 0, 'no'],

									        [0, 1, 'no'],

									        [0, 1, 'no'],

									        [0,0,'maybe']]

									  labels = ['no surfaceing', 'flippers']

									  return dataSet, labels

									# 计算香农熵

									def calcShannonEnt(dataSet):

									  numEntries = len(dataSet)

									  labelCounts = {}

									  for feaVec in dataSet:

									    currentLabel = feaVec[-1]

									    if currentLabel not in labelCounts:

									      labelCounts[currentLabel] = 0

									    labelCounts[currentLabel] += 1

									  shannonEnt = 0.0

									  for key in labelCounts:

									    prob = float(labelCounts[key]) / numEntries

									    shannonEnt -= prob * log(prob, 2)

									  return shannonEnt

									def splitDataSet(dataSet, axis, value):

									  retDataSet = []

									  for featVec in dataSet:

									    if featVec[axis] == value:

									      reducedFeatVec = featVec[:axis]

									      reducedFeatVec.extend(featVec[axis + 1:])

									      retDataSet.append(reducedFeatVec)

									  return retDataSet

									def chooseBestFeatureToSplit(dataSet):

									  numFeatures = len(dataSet[0]) - 1 # 因为数据集的最后一项是标签

									  baseEntropy = calcShannonEnt(dataSet)

									  bestInfoGain = 0.0

									  bestFeature = -1

									  for i in range(numFeatures):

									    featList = [example[i] for example in dataSet]

									    uniqueVals = set(featList)

									    newEntropy = 0.0

									    for value in uniqueVals:

									      subDataSet = splitDataSet(dataSet, i, value)

									      prob = len(subDataSet) / float(len(dataSet))

									      newEntropy += prob * calcShannonEnt(subDataSet)

									    infoGain = baseEntropy - newEntropy

									    if infoGain > bestInfoGain:

									      bestInfoGain = infoGain

									      bestFeature = i

									  return bestFeature

									# 因为我们递归构建决策树是根据属性的消耗进行计算的，所以可能会存在最后属性用完了，但是分类

									# 还是没有算完，这时候就会采用多数表决的方式计算节点分类

									def majorityCnt(classList):

									  classCount = {}

									  for vote in classList:

									    if vote not in classCount.keys():

									      classCount[vote] = 0

									    classCount[vote] += 1

									  return max(classCount)

									def createTree(dataSet, labels):

									  classList = [example[-1] for example in dataSet]

									  if classList.count(classList[0]) == len(classList): # 类别相同则停止划分

									    return classList[0]

									  if len(dataSet[0]) == 1: # 所有特征已经用完

									    return majorityCnt(classList)

									  bestFeat = chooseBestFeatureToSplit(dataSet)

									  bestFeatLabel = labels[bestFeat]

									  myTree = {bestFeatLabel: {}}

									  del (labels[bestFeat])

									  featValues = [example[bestFeat] for example in dataSet]

									  uniqueVals = set(featValues)

									  for value in uniqueVals:

									    subLabels = labels[:] # 为了不改变原始列表的内容复制了一下

									    myTree[bestFeatLabel][value] = createTree(splitDataSet(dataSet,

									                                bestFeat, value), subLabels)

									  return myTree

									def main():

									  data, label = createDataSet()

									  t1 = time.clock()

									  myTree = createTree(data, label)

									  t2 = time.clock()

									  print myTree

									  print 'execute for ', t2 - t1

									if __name__ == '__main__':

									  main()

运行结果如下：

1 2	`{'no surfaceing': {0: {'flippers': {0:` `'maybe',` `1:` `'no'}},` `1: {'flippers': {0:` `'no',` `1:` `'yes'}}}}` `execute` `for` `0.0103958394532`