版权声明:本文为博主原创文章,未经博主允许不得转载。
本文基于python逐步实现Decision Tree(决策树),分为以下几个步骤:
- 加载数据集
- 熵的计算
- 根据最佳分割feature进行数据分割
- 根据最大信息增益选择最佳分割feature
- 递归构建决策树
- 样本分类
关于决策树的理论方面本文几乎不讲,详情请google keywords:“决策树 信息增益 熵”
将分别体现于代码。
本文只建一个.py文件,所有代码都在这个py里
1.加载数据集
我们选用UCI经典Iris为例
Brief of IRIS:
Data Set Characteristics: |
Multivariate |
Number of Instances: |
150 |
Area: |
Life |
Attribute Characteristics: |
Real |
Number of Attributes: |
4 |
Date Donated |
1988-07-01 |
Associated Tasks: |
Classification |
Missing Values? |
No |
Number of Web Hits: |
533125 |
Code:
- from numpy import *
- #load "iris.data" to workspace
- traindata = loadtxt("D:\ZJU_Projects\machine learning\ML_Action\Dataset\Iris.data",delimiter = ',',usecols = (0,1,2,3),dtype = float)
- trainlabel = loadtxt("D:\ZJU_Projects\machine learning\ML_Action\Dataset\Iris.data",delimiter = ',',usecols = (range(4,5)),dtype = str)
- feaname = ["#0","#1","#2","#3"] # feature names of the 4 attributes (features)
Result:
左图为实际数据集,四个离散型feature,一个label表示类别(有Iris-setosa, Iris-versicolor,Iris-virginica 三个类)
2. 熵的计算
entropy是香农提出来的(信息论大牛),定义见wiki
注意这里的entropy是H(C|X=xi)而非H(C|X), H(C|X)的计算见第下一个点,还要乘以概率加和
Code:
- from math import log
- def calentropy(label):
- n = label.size # the number of samples
- #print n
- count = {} #create dictionary "count"
- for curlabel in label:
- if curlabel not in count.keys():
- count[curlabel] = 0
- count[curlabel] += 1
- entropy = 0
- #print count
- for key in count:
- pxi = float(count[key])/n #notice transfering to float first
- entropy -= pxi*log(pxi,2)
- return entropy
- #testcode:
- #x = calentropy(trainlabel)
Result:
3. 根据最佳分割feature进行数据分割
假定我们已经得到了最佳分割feature,在这里进行分割(最佳feature为splitfea_idx)
第二个函数idx2data是根据splitdata得到的分割数据的两个index集合返回datal (samples less than pivot), datag(samples greater than pivot), labell, labelg。 这里我们根据所选特征的平均值作为pivot
- #split the dataset according to label "splitfea_idx"
- def splitdata(oridata,splitfea_idx):
- arg = args[splitfea_idx] #get the average over all dimensions
- idx_less = [] #create new list including data with feature less than pivot
- idx_greater = [] #includes entries with feature greater than pivot
- n = len(oridata)
- for idx in range(n):
- d = oridata[idx]
- if d[splitfea_idx] < arg:
- #add the newentry into newdata_less set
- idx_less.append(idx)
- else:
- idx_greater.append(idx)
- return idx_less,idx_greater
- #testcode:2
- #idx_less,idx_greater = splitdata(traindata,2)
- #give the data and labels according to index
- def idx2data(oridata,label,splitidx,fea_idx):
- idxl = splitidx[0] #split_less_indices
- idxg = splitidx[1] #split_greater_indices
- datal = []
- datag = []
- labell = []
- labelg = []
- for i in idxl:
- datal.append(append(oridata[i][:fea_idx],oridata[i][fea_idx+1:]))
- for i in idxg:
- datag.append(append(oridata[i][:fea_idx],oridata[i][fea_idx+1:]))
- labell = label[idxl]
- labelg = label[idxg]
- return datal,datag,labell,labelg
这里args是参数,决定分裂节点的阈值(每个参数对应一个feature,大于该值分到>branch,小于该值分到<branch),我们可以定义如下:
- args = mean(traindata,axis = 0)
测试:按特征2进行分类,得到的less和greater set of indices分别为:
也就是按args[2]进行样本集分割,<和>args[2]的branch分别有57和93个样本。
4. 根据最大信息增益选择最佳分割feature
信息增益为代码中的info_gain, 注释中是熵的计算
- #select the best branch to split
- def choosebest_splitnode(oridata,label):
- n_fea = len(oridata[0])
- n = len(label)
- base_entropy = calentropy(label)
- best_gain = -1
- for fea_i in range(n_fea): #calculate entropy under each splitting feature
- cur_entropy = 0
- idxset_less,idxset_greater = splitdata(oridata,fea_i)
- prob_less = float(len(idxset_less))/n
- prob_greater = float(len(idxset_greater))/n
- #entropy(value|X) = \sum{p(xi)*entropy(value|X=xi)}
- cur_entropy += prob_less*calentropy(label[idxset_less])
- cur_entropy += prob_greater * calentropy(label[idxset_greater])
- info_gain = base_entropy - cur_entropy #notice gain is before minus after
- if(info_gain>best_gain):
- best_gain = info_gain
- best_idx = fea_i
- return best_idx
- #testcode:
- #x = choosebest_splitnode(traindata,trainlabel)
这里的测试针对所有数据,分裂一次选择哪个特征呢?
5. 递归构建决策树
详见code注释,buildtree递归地构建树。
递归终止条件:
①该branch内没有样本(subset为空) or
②分割出的所有样本属于同一类 or
③由于每次分割消耗一个feature,当没有feature的时候停止递归,返回当前样本集中大多数sample的label
- #create the decision tree based on information gain
- def buildtree(oridata, label):
- if label.size==0: #if no samples belong to this branch
- return "NULL"
- listlabel = label.tolist()
- #stop when all samples in this subset belongs to one class
- if listlabel.count(label[0])==label.size:
- return label[0]
- #return the majority of samples' label in this subset if no extra features avaliable
- if len(feanamecopy)==0:
- cnt = {}
- for cur_l in label:
- if cur_l not in cnt.keys():
- cnt[cur_l] = 0
- cnt[cur_l] += 1
- maxx = -1
- for keys in cnt:
- if maxx < cnt[keys]:
- maxx = cnt[keys]
- maxkey = keys
- return maxkey
- bestsplit_fea = choosebest_splitnode(oridata,label) #get the best splitting feature
- print bestsplit_fea,len(oridata[0])
- cur_feaname = feanamecopy[bestsplit_fea] # add the feature name to dictionary
- print cur_feaname
- nodedict = {cur_feaname:{}}
- del(feanamecopy[bestsplit_fea]) #delete current feature from feaname
- split_idx = splitdata(oridata,bestsplit_fea) #split_idx: the split index for both less and greater
- data_less,data_greater,label_less,label_greater = idx2data(oridata,label,split_idx,bestsplit_fea)
- #build the tree recursively, the left and right tree are the "<" and ">" branch, respectively
- nodedict[cur_feaname]["<"] = buildtree(data_less,label_less)
- nodedict[cur_feaname][">"] = buildtree(data_greater,label_greater)
- return nodedict
- #testcode:
- #mytree = buildtree(traindata,trainlabel)
- #print mytree
Result:
mytree就是我们的结果,#1表示当前使用第一个feature做分割,'<'和'>'分别对应less 和 greater的数据。
6. 样本分类
根据构建出的mytree进行分类,递归走分支
- #classify a new sample
- def classify(mytree,testdata):
- if type(mytree).__name__ != 'dict':
- return mytree
- fea_name = mytree.keys()[0] #get the name of first feature
- fea_idx = feaname.index(fea_name) #the index of feature 'fea_name'
- val = testdata[fea_idx]
- nextbranch = mytree[fea_name]
- #judge the current value > or < the pivot (average)
- if val>args[fea_idx]:
- nextbranch = nextbranch[">"]
- else:
- nextbranch = nextbranch["<"]
- return classify(nextbranch,testdata)
- #testcode
- tt = traindata[0]
- x = classify(mytree,tt)
- print x
Result:
为了验证代码准确性,我们换一下args参数,把它们都设成0(很小)
args = [0,0,0,0]
建树和分类的结果如下:
可见没有小于pivot(0)的项,于是dict中每个<的key对应的value都为空。
本文中全部代码下载:决策树python实现
Reference: Machine Learning in Action
from: http://blog.****.net/abcjennifer/article/details/20905311