1+ # -*- coding: utf-8 -*-
2+ """
3+ Created on Wed Feb 24 08:57:10 2016
4+ Adaboost algorithm
5+ @author: liudiwei
6+ """
7+ import numpy as np
8+
9+
10+ class AdaboostClassifier (object ):
11+
12+ def __init__ (self , max_iter = 100 , numSteps = 10 , num_iter = 50 ):
13+ self ._max_iter = max_iter
14+ self ._numSteps = numSteps
15+ self ._num_iter = num_iter
16+
17+ def _stumpClassify (self , X , axis , threshVal , threshIneq ):
18+ """
19+ 功能：一个单层决策树，通过阈值比较对数据集进行分类
20+ 返回：
21+ """
22+ retArray = np .ones ((np .shape (X )[0 ], 1 )) #生成一个m*1的矩阵，全1
23+ if threshIneq == 'lt' : #设定规则为"lt"，小于阈值则为-1.0
24+ retArray [X [:, axis ] <= threshVal ] = - 1.0
25+ else :
26+ retArray [X [:, axis ] > threshVal ] = - 1.0
27+ return retArray
28+
29+ def _buildStump (self , X , y , D ):
30+ """
31+ 输入：X, y和初始化的权重值D
32+ 功能：找到最佳的单层决策树
33+ 返回：一个字典(dim, ineq, thresh)、错误率、基于该特征和阈值下的类别估计值
34+ """
35+ dataMat = np .mat (X )
36+ labelMat = np .mat (y ).T #类标签转置
37+ m , n = np .shape (dataMat ) #获得矩阵的维数
38+ bestStump = {}
39+ bestClasEst = np .mat (np .zeros ((m ,1 )))
40+ minError = np .inf #初始化最小误差，令其为最大值
41+ for i in range (n ):#遍历所有的维数
42+ minVal = dataMat [:,i ].min ()
43+ maxVal = dataMat [:,i ].max ()
44+ stepSize = (maxVal - minVal )/ self ._numSteps
45+ for j in range (- 1 , int (self ._numSteps )+ 1 ):
46+ for inequal in ['lt' , 'gt' ]:
47+ threshVal = (minVal + float (j )* stepSize )
48+ predVals = self ._stumpClassify (dataMat ,i ,threshVal ,inequal )
49+ errArr = np .mat (np .ones ((m ,1 )))
50+ errArr [predVals == labelMat ] = 0 #统计误分类的个数
51+ weightedError = D .T * errArr #计算加权错误率
52+ if weightedError < minError :
53+ minError = weightedError
54+ bestClasEst = predVals .copy ()
55+ bestStump ['dim' ] = i
56+ bestStump ['thresh' ] = threshVal
57+ bestStump ['ineq' ] = inequal
58+ return bestStump , minError , bestClasEst
59+
60+
61+ def fit (self , X , y ):
62+ """
63+ 输入：X，y，迭代次数
64+ 输出：弱分类器weakClassArr（字典）
65+ """
66+ weakClassArr = []
67+ m = np .shape (X )[0 ]
68+ D = np .mat (np .ones ((m , 1 ))/ m ) #初始化权重值都相等，为1/m
69+ aggClassEst = np .mat (np .zeros ((m , 1 ))) #样本点类别估计累计值
70+ for i in range (self ._num_iter ):
71+ bestStump , error , classEst = self ._buildStump (X , y , D )
72+ alpha = float (0.5 * np .log ((1.0 - error ) / max (error , 1e-16 )))
73+ bestStump ['alpha' ] = alpha
74+ weakClassArr .append (bestStump )
75+ expon = np .multiply (- 1 * alpha * np .mat (y ).T , classEst )
76+ D = np .multiply (D , np .exp (expon ))
77+ D = D / D .sum ()
78+ #计算所有分类器的训练误差，如果为0则提前结束循环
79+ aggClassEst += alpha * classEst
80+ aggErrors = np .multiply (\
81+ np .sign (aggClassEst ) != np .mat (y ).T ,np .ones ((m ,1 )))
82+ errorRate = aggErrors .sum ()/ m
83+ #print "total error is: ", errorRate
84+ if errorRate == 0.0 :
85+ break
86+ return weakClassArr
87+
88+ def predict (self , test_X , classifierArr ):
89+ """
90+ 输入: 测试集和分类器
91+ 输出：分类结果（二分类）
92+ """
93+ dataMat = np .mat (test_X )
94+ m = np .shape (dataMat )[0 ]
95+ aggClassEst = np .mat (np .zeros ((m ,1 )))
96+ for i in range (len (classifierArr )):
97+ classEst = self ._stumpClassify (dataMat ,\
98+ classifierArr [i ]['dim' ],\
99+ classifierArr [i ]['thresh' ],\
100+ classifierArr [i ]['ineq' ])
101+ aggClassEst += classifierArr [i ]['alpha' ] * classEst
102+ #print aggClassEst
103+ return np .sign (aggClassEst )
0 commit comments