|
Logistic回归
算法优缺点:
1.计算代价不高,易于理解和实现
2.容易欠拟合,分类精度可能不高
3.适用数据类型:数值型和标称型
算法思想:
- 其实就我的理解来说,logistic回归实际上就是加了个sigmoid函数的线性回归,这个sigmoid函数的好处就在于,将结果归到了0到1这个区间里面了,并且sigmoid(0)=0.5,也就是说里面的线性部分的结果大于零小于零就可以直接计算到了。这里的求解方式是梯度上升法,具体我就不扯了,最推荐的资料还是Ng的视频,那里面的梯度下降就是啦,只不过一个是梯度上升的方向一个是下降的方向,做法什么的都一样。
- 而梯度上升(准确的说叫做“批梯度上升”)的一个缺点就是计算量太大了,每一次迭代都需要把所有的数据算一遍,这样一旦训练集大了之后,那么计算量将非常大,所以这里后面还提出了随机梯度下降,思想就是每次只是根据一个data进行修正。这样得到的最终的结果可能会有所偏差但是速度却提高了很多,而且优化之后的偏差还是很小的。随机梯度上升的另一个好处是这是一个在线算法,可以根据新数据的到来不断处理
函数:
loadDataSet()
创建数据集,这里的数据集就是在一个文件中,这里面有三行,分别是两个特征和一个标签,但是我们在读出的时候还加了X0这个属性
sigmoid(inX)
sigmoid函数的计算,这个函数长这样的,基本坐标大点就和阶跃函数很像了
gradAscend(dataMatIn, classLabels)
梯度上升算法的实现,里面用到了numpy的数组,并且设定了迭代次数500次,然后为了计算速度都采取了矩阵计算,计算的过程中的公式大概是:w= w+alpha*(y-h)x(一直懒得写公式,见谅。。。)
gradAscendWithDraw(dataMatIn, classLabels)
上面的函数加强版,增加了一个weight跟着迭代次数的变化曲线
stocGradAscent0(dataMatrix, classLabels)
这里为了加快速度用来随机梯度上升,即每次根据一组数据调整(额,好吧,这个际没有随机因为那是线面那个函数)
stocGradAscentWithDraw0(dataMatrix, classLabels)
上面的函数加强版,增加了一个weight跟着迭代次数的变化曲线
stocGradAscent1(dataMatrix, classLabels, numIter=150)
这就真的开始随机了,随机的主要好处是减少了周期性的波动了。另外这里还加入了alpha的值随迭代变化,这样可以让alpha的值不断的变化,但是都不会减小到0。
stocGradAscentWithDraw1(dataMatrix, classLabels, numIter=150)
上面的函数加强版,增加了一个weight跟着迭代次数的变化曲线
plotBestFit(wei)
根据计算的weight值画出拟合的线,直观观察效果
运行效果分析:
1、梯度上升:
迭代变化趋势
分类结果:
2、随机梯度上升版本1
迭代变化趋势
分类结果:
这个速度虽然快了很多但是效果不太理想啊。不过这个计算量那么少,我们如果把这个迭代200次肯定不一样了,效果如下
果然好多了
3、随机梯度上升版本2
迭代变化趋势
分类结果:
恩,就是这样啦,效果还是不错的啦。代码的画图部分写的有点烂,见谅啦
1 #coding=utf-8
2 from numpy import *
3
4 def loadDataSet():
5 dataMat = []
6 labelMat = []
7 fr = open('testSet.txt')
8 for line in fr.readlines():
9 lineArr = line.strip().split()
10 dataMat.append([1.0, float(lineArr[0]), float(lineArr[1])])
11 labelMat.append(int(lineArr[2]))
12 return dataMat, labelMat
13
14 def sigmoid(inX):
15 return 1.0/(1+exp(-inX))
16
17 def gradAscend(dataMatIn, classLabels):
18 dataMatrix = mat(dataMatIn)
19 labelMat = mat(classLabels).transpose()
20 m,n = shape(dataMatrix)
21 alpha = 0.001
22 maxCycle = 500
23 weight = ones((n,1))
24 for k in range(maxCycle):
25 h = sigmoid(dataMatrix*weight)
26 error = labelMat - h
27 weight += alpha * dataMatrix.transpose() * error
28 #plotBestFit(weight)
29 return weight
30
31 def gradAscendWithDraw(dataMatIn, classLabels):
32 import matplotlib.pyplot as plt
33 fig = plt.figure()
34 ax = fig.add_subplot(311,ylabel='x0')
35 bx = fig.add_subplot(312,ylabel='x1')
36 cx = fig.add_subplot(313,ylabel='x2')
37 dataMatrix = mat(dataMatIn)
38 labelMat = mat(classLabels).transpose()
39 m,n = shape(dataMatrix)
40 alpha = 0.001
41 maxCycle = 500
42 weight = ones((n,1))
43 wei1 = []
44 wei2 = []
45 wei3 = []
46 for k in range(maxCycle):
47 h = sigmoid(dataMatrix*weight)
48 error = labelMat - h
49 weight += alpha * dataMatrix.transpose() * error
50 wei1.extend(weight[0])
51 wei2.extend(weight[1])
52 wei3.extend(weight[2])
53 ax.plot(range(maxCycle), wei1)
54 bx.plot(range(maxCycle), wei2)
55 cx.plot(range(maxCycle), wei3)
56 plt.xlabel('iter_num')
57 plt.show()
58 return weight
59
60 def stocGradAscent0(dataMatrix, classLabels):
61 m,n = shape(dataMatrix)
62
63 alpha = 0.001
64 weight = ones(n)
65 for i in range(m):
66 h = sigmoid(sum(dataMatrix*weight))
67 error = classLabels - h
68 weight = weight + alpha * error * dataMatrix
69 return weight
70
71 def stocGradAscentWithDraw0(dataMatrix, classLabels):
72 import matplotlib.pyplot as plt
73 fig = plt.figure()
74 ax = fig.add_subplot(311,ylabel='x0')
75 bx = fig.add_subplot(312,ylabel='x1')
76 cx = fig.add_subplot(313,ylabel='x2')
77 m,n = shape(dataMatrix)
78
79 alpha = 0.001
80 weight = ones(n)
81 wei1 = array([])
82 wei2 = array([])
83 wei3 = array([])
84 numIter = 200
85 for j in range(numIter):
86 for i in range(m):
87 h = sigmoid(sum(dataMatrix*weight))
88 error = classLabels - h
89 weight = weight + alpha * error * dataMatrix
90 wei1 =append(wei1, weight[0])
91 wei2 =append(wei2, weight[1])
92 wei3 =append(wei3, weight[2])
93 ax.plot(array(range(m*numIter)), wei1)
94 bx.plot(array(range(m*numIter)), wei2)
95 cx.plot(array(range(m*numIter)), wei3)
96 plt.xlabel('iter_num')
97 plt.show()
98 return weight
99
100 def stocGradAscent1(dataMatrix, classLabels, numIter=150):
101 m,n = shape(dataMatrix)
102
103 #alpha = 0.001
104 weight = ones(n)
105 for j in range(numIter):
106 dataIndex = range(m)
107 for i in range(m):
108 alpha = 4/ (1.0+j+i) +0.01
109 randIndex = int(random.uniform(0,len(dataIndex)))
110 h = sigmoid(sum(dataMatrix[randIndex]*weight))
111 error = classLabels[randIndex] - h
112 weight = weight + alpha * error * dataMatrix[randIndex]
113 del(dataIndex[randIndex])
114 return weight
115
116 def stocGradAscentWithDraw1(dataMatrix, classLabels, numIter=150):
117 import matplotlib.pyplot as plt
118 fig = plt.figure()
119 ax = fig.add_subplot(311,ylabel='x0')
120 bx = fig.add_subplot(312,ylabel='x1')
121 cx = fig.add_subplot(313,ylabel='x2')
122 m,n = shape(dataMatrix)
123
124 #alpha = 0.001
125 weight = ones(n)
126 wei1 = array([])
127 wei2 = array([])
128 wei3 = array([])
129 for j in range(numIter):
130 dataIndex = range(m)
131 for i in range(m):
132 alpha = 4/ (1.0+j+i) +0.01
133 randIndex = int(random.uniform(0,len(dataIndex)))
134 h = sigmoid(sum(dataMatrix[randIndex]*weight))
135 error = classLabels[randIndex] - h
136 weight = weight + alpha * error * dataMatrix[randIndex]
137 del(dataIndex[randIndex])
138 wei1 =append(wei1, weight[0])
139 wei2 =append(wei2, weight[1])
140 wei3 =append(wei3, weight[2])
141 ax.plot(array(range(len(wei1))), wei1)
142 bx.plot(array(range(len(wei2))), wei2)
143 cx.plot(array(range(len(wei2))), wei3)
144 plt.xlabel('iter_num')
145 plt.show()
146 return weight
147
148 def plotBestFit(wei):
149 import matplotlib.pyplot as plt
150 weight = wei
151 dataMat,labelMat = loadDataSet()
152 dataArr = array(dataMat)
153 n = shape(dataArr)[0]
154 xcord1 = []
155 ycord1 = []
156 xcord2 = []
157 ycord2 = []
158 for i in range(n):
159 if int(labelMat) == 1:
160 xcord1.append(dataArr[i,1])
161 ycord1.append(dataArr[i,2])
162 else:
163 xcord2.append(dataArr[i,1])
164 ycord2.append(dataArr[i,2])
165 fig = plt.figure()
166 ax = fig.add_subplot(111)
167 ax.scatter(xcord1, ycord1, s=30, c='red', marker='s')
168 ax.scatter(xcord2, ycord2, s=30, c='green')
169 x = arange(-3.0, 3.0, 0.1)
170 y = (-weight[0] - weight[1]*x)/weight[2]
171 ax.plot(x,y)
172 plt.xlabel('X1')
173 plt.ylabel('X2')
174 plt.show()
175
176 def main():
177 dataArr,labelMat = loadDataSet()
178 #w = gradAscendWithDraw(dataArr,labelMat)
179 w = stocGradAscentWithDraw0(array(dataArr),labelMat)
180 plotBestFit(w)
181
182 if __name__ == '__main__':
183 main()
机器学习笔记索引
来自为知笔记(Wiz) |
|