Fisher准则线性分类器的Python实现

Fisher准则线性分类器的Python实现

本节内容：本节内容是根据上学期所上的模式识别课程的作业整理而来，第二道题目是线性分类器设计，数据集是Iris(鸢尾花的数据集)，根据前一题的Kmeans聚类得出的结果，分成训练集与测试集，进行比较。

---

选取的训练集与测试集

1. 训练集：(选取上一题中的第一种结果:每一类大小都是l=0.67*len(dataset[i])，前l个数据)
   第一类（33个）：[[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
   22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32],
   第二类（25个）：[52, 77, 100, 102, 103, 104, 105, 107, 108, 109, 110, 111, 112, 115, 116, 117, 118, 120, 122, 124, 125, 128, 129, 130, 131],
   第三类（41个）：[50, 51, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92]]
2. 测试集
   第一类(17个)：[[33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
   第二类(13个）：[132, 134, 135, 136, 137, 139, 140, 141, 143, 144, 145, 147, 148],
   第三类(21个)： [93, 94, 95, 96, 97, 98, 99, 101, 106, 113, 114, 119, 121, 123, 126, 127, 133, 138, 142, 146, 149]]

---

分类决策与分类器

1. 分类决策：一对一
2. 分类器：Fisher线性分类器
   $g_{12}(x)=W_{12}^Tx-W_{012}, g_{13}(x)=W_{13}^Tx-W_{013}, g_{23}(x)=W_{23}^Tx-W_{023}, g_{21}(x)= -g_{12}(x)，g_{31}(x)= -g_{13}(x)，g_{32}(x)= -g_{23}(x)$
3. 相关参数
   类间离散度矩阵：$S_{12}=(m_1-m_2)(m_1-m_2)^T$
   类内离散度矩阵：$S_w=\sum_{x_i \in \omega_i}(x_i-m_1)(x_i-m_1)^T$
   最优投影方向：$W^*=S_w^{-1}(m_1-m_2)$
4. 判别函数
   第一类：g12(x)>0 and g13(x)>0
   第二类：g12(x)<0 and g23(x)>0
   第三类：g13(x)<0 and g23(x)<0
   拒识情况：其他
   在本题中，以上分类器分别为：

g12(x) =(W12.T)*(X.T)-W012，W12=la.inv(sw12)*((u01-u02).T)
W012=(l1*(W12.T)*(u01.T)+l2*(W12.T)*(u02.T))/(l1+l2)
g13(x) =(W13.T)*(X.T)-W013，  W13=la.inv(sw13)*((u01-u03).T)
W013=(l1*(W13.T)*(u01.T)+l3*(W13.T)*(u03.T))/(l1+l3)
g23(x)= (W23.T)*(X.T)-W023，W23=la.inv(sw23)*(u02-u03).T
W023=(l2*(W23.T)*(u02.T)+l3*(W23.T)*(u03.T))/(l3+l2)

---

代码

# coding=gbk
#python edition: Python3.4.1,2014,10,17
import numpy as np
from numpy import linalg as la

def read_points():
    dataset=[]
    with open('Iris.txt','r') as file:
        for line in file:
            if line =='\n':
                continue
            dataset.append(list(map(float,line.split(' '))))
        file.close()
        return  dataset

def generate_traineddata():
    arr=[[] for i in range(3)]
    with open('setbase.txt','r') as file:
        index=0
        for line in file:
            if line=='\n' :
                continue
            elif line[0]=='C':
                index=int(line[-2])-1
                continue
            arr[index].append(int(line))
        file.close()
    train=[[] for i in range(3)]
    test=[[] for i in range(3)]
    for i in range(len(arr)):
        tr=int(0.67* len(arr[i]))
        train[i]=arr[i][:tr]
        test[i]=arr[i][tr:]
    f1=open('trained.txt','w')
    f2=open('tested.txt','w')
    print(train,end='\n')
    print(test,end='\n')
    for i in range(3):
        for j in train[i]:
            f1.write("%d\n"%j)
        f1.write('\n')
        for k in test[i]:
            f2.write("%d\n"%k)
        f2.write('\n')
    f1.close()
    f2.close()
    return train,test

def createMatrix(train,test,dataset):
    trainmat=[[] for i in range(3)]
    testmat=[[] for i in range(3)]
    for i in range(3):
        for j in train[i]:
            trainmat[i].append(dataset[j])
        for k in test[i]:
            testmat[i].append(dataset[k])
    return   trainmat,testmat

def classify(trainmat,testmat,test):
    #求三类训练集的均值向量
    tr1,tr2,tr3=np.mat(trainmat[0]),np.mat(trainmat[1]),np.mat(trainmat[2])
    te=[[] for i in range(3)]
    te[0],te[1],te[2]=np.mat(testmat[0]),np.mat(testmat[1]),np.mat(testmat[2])
    u01=np.mean(tr1,axis=0)
    u02=np.mean(tr2,axis=0)
    u03=np.mean(tr3,axis=0)
    #获得矩阵长度
    l1,l2,l3=len(trainmat[0]),len(trainmat[1]),len(trainmat[2])
    #求三类训练集的类内离散度矩阵
    s1,s2,s3=0,0,0
    for i in range(l1):
        s1=s1+(tr1[i]-u01).T*(tr1[i]-u01)
    for i in range(l2):
        s2=s2+ (tr2[i]-u02).T*(tr2[i]-u02)
    for i in range(l3):
        s3=s3+ (tr3[i]-u03).T*(tr3[i]-u03)
    #总类内离散度矩阵
    sw12,sw13,sw23=s1+s2,s1+s3, s2+s3
    #求向量W*与边界
    W12=la.inv(sw12)*((u01-u02).T)
    W012=(l1*(W12.T)*(u01.T)+l2*(W12.T)*(u02.T))/(l1+l2)
    W13=la.inv(sw13)*((u01-u03).T)
    W013=(l1*(W13.T)*(u01.T)+l3*(W13.T)*(u03.T))/(l1+l3)
    W23=la.inv(sw23)*(u02-u03).T
    W023=(l2*(W23.T)*(u02.T)+l3*(W23.T)*(u03.T))/(l3+l2)
    result=[[] for i in range(4)]
    for i in range(3):
        testset=te[i]
        count=0
        for X in testset:
            if ((W12.T)*(X.T)-W012>0) and   ((W13.T)*(X.T)-W013>0):
                result[0].append(test[i][count])
            elif  ((W12.T)*(X.T)-W012<0) and   ((W23.T)*(X.T)-W023>0):
                result[1].append(test[i][count])
            elif    ((W13.T)*(X.T)-W013<0) and   ((W23.T)*(X.T)-W023<0):
                result[2].append(test[i][count])
            else:
                result[3].append(test[i][count])
            count=count+1
    str1='类内离散度矩阵：'
    str2='投影方向：'
    str3='边界点：'
    str4='Fisher得出的分类结果：'
    fisher=open('fisher.txt','w')
    print(str1,'s1:',end='\n',file=fisher)
    print(s1,end='\n',file=fisher)
    print(str1,'s2:',end='\n',file=fisher)
    print(s2,end='\n',file=fisher)
    print(str1,'s3:',end='\n',file=fisher)
    print(s3,end='\n',file=fisher)
    print(str2,'\n','W12:\n',W12,end='\n',file=fisher)
    print('W13:\n',W13,end='\n',file=fisher)
    print('W23:\n',W23,end='\n',file=fisher)
    print(str3,'\n','W012:',W012,end='\n' ,file=fisher)
    print('W013:',W013,end='\n' ,file=fisher)
    print('W023:',W023,end='\n' ,file=fisher)
    print(str4,end='\n' ,file=fisher)
    for i in range(4):
        print('第%d类'%(i+1),result[i],end='\n' ,file=fisher)
    fisher.close()
    return    result


def main():
    dataset=read_points()
    train,test= generate_traineddata()
    trainmat,testmat=createMatrix(train,test,dataset)
    result= classify(trainmat,testmat,test)
    print(result)

if __name__=='__main__':
    main()

---

测试集上的结果

1. 实验中测得的参数：
   类内离散度矩阵： s1:
   [[ 4.23636364 3.10090909 0.50545455 0.50545455]
   [ 3.10090909 4.10060606 0.02030303 0.47363636]
   [ 0.50545455 0.02030303 1.01515152 0.16181818]
   [ 0.50545455 0.47363636 0.16181818 0.34181818]]
   类内离散度矩阵： s2:
   [[ 6.12 0.948 5.132 0.046 ]
   [ 0.948 2.5144 0.4916 0.6928]
   [ 5.132 0.4916 7.0024 1.2592]
   [ 0.046 0.6928 1.2592 1.6136]]
   类内离散度矩阵： s3:
   [[ 9.60097561 3.05414634 5.14512195 1.73707317]
   [ 3.05414634 4.27512195 2.67926829 1.69756098]
   [ 5.14512195 2.67926829 7.34439024 2.53463415]
   [ 1.73707317 1.69756098 2.53463415 1.56878049]]
   投影方向：
   W12: [[ 0.00467303], [ 0.21663412], [-0.43497883], [-0.71989691]]T
   W13: [[-0.01511481], [ 0.3200074 ], [-0.25327497], [-0.55700484]]T
   W23: [[ 0.02024091], [-0.05645396], [ 0.0580067 ], [ 0.17693648]]T
   边界点：
   W012: [[-1.44922042]], W013: [[-0.34557374]], W023: [[ 0.53145747]]
2. 分类结果（无拒识情况，错误的为斜黑体）
   第1类 [33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49]
   第2类 [132, 134, 135, 136, 137, 139, 140, 141, 143, 144, 145, 147, 148, 101, 113, 114, 119, 121, 123, 126, 127, 138, 142, 146, 149]
   本身为第三类，却分到了第二类
   第3类 [93, 94, 95, 96, 97, 98, 99, 106, 133]
3. 正确率
   39/51*100%=76.47%