logistic回归是统计学习中经典的分类方法,他属于对数线性模型。本博文主要给出logistic的c++实现,具体理论请读者自行google。
本文用到的数据集来自于一个医学网站,具体出处不记得了(非常歉意)。数据的格式如下:
10009 1 0 0 1 0 1
10025 0 0 1 2 0 0
20035 0 0 1 0 0 1
20053 1 0 0 0 0 0
30627 1 0 1 2 0 0
30648 2 0 0 0 1 0
每行有7个列值,第一列是一个ID号,在具体操作中,忽略该列。之后的5列,每一个都表示一个特征的取值;最后一列是分类标记(0或1)。
在具体实现时,将分隔数据为训练数据和测试数据,并保存到文件中,文件组织形式如下:
其中testdata.txt,保存测试数据;traindata.txt保存训练数据;logistic.cpp是代码源文件。三个文件保存在同一目录下。
实现代码如下:
/*********
logistic回归(c++) by 姜富春
**********/
#include<iostream>
#include<fstream>
#include<vector>
#include<sstream>
#include<cmath>
using namespace std;
struct Data{
vector<int> features;
int cls;
Data(vector<int> f,int c):features(f),cls(c){
}
};
struct Param{
vector<double> w;
double d;
Param(vector<double> w1,double d1):w(w1),d(d1){};
Param():w(vector<double>()),d(0.0){}
};
class Logistic{
public:
Logistic(){
//载入traindata文件构造dataSet;
loadDataSet(dataSet);
//初始化Param,w的长度与数据特征的长度相同,初值为0.0。d的初值也为0.0
vector<double> pw(dataSet[0].features.size(),0.0);
Param pt(pw,0.0);
param=pt;
};
void loadDataSet(vector<Data>& ds,string dataFile="./traindata.txt"){
ifstream fin(dataFile.c_str());
if(!fin){
cout<<"文件打开失败"<<endl;
exit(0);
}
while(fin){
string line;
getline(fin,line);
if(line.size()>3){
stringstream sin(line);
int t;
sin>>t;
vector<int> fea;
while(sin){
char c=sin.peek();
if(int(c)!=-1){
sin>>t;
fea.push_back(t);
}
}
int cl=fea.back();
fea.pop_back();
ds.push_back(Data(fea,cl));
}
}
}
void displayDataSet(){
for(int i=0;i<dataSet.size();i++){
for(int j=0;j<dataSet[i].features.size();j++){
cout<<dataSet[i].features[j]<<" ";
}
cout<<" 分类:"<<dataSet[i].cls;
cout<<endl;
}
}
void logisticRegression(){
//由目标函数为最大似然,因此最终求得的是目标函数的最大值,
//因此迭代过程是梯度上升,而非梯度下降
double lamda=0.1;//梯度下降的步长
double delta=0.0001;//结束迭代的阈值
//目标函数的值
double objLw=Lw(param);
//cout<<objLw<<endl;
Param tpa(param.w,param.d);
gradient(lamda);
double newObjLw=Lw(param);
int iter=0;
cout<<"初始:"<<endl;
displayIterProcess(iter,objLw,newObjLw,1);
while(fabs(newObjLw-objLw)>delta||!samewb(tpa,param,delta)){
objLw=newObjLw;
tpa=Param(param.w,param.d);
gradient(lamda);
newObjLw=Lw(param);
++iter;
displayIterProcess(iter,objLw,newObjLw,5);
}
cout<<"迭代结束共迭代"<<iter<<"步"<<endl;
displayIterProcess(iter,objLw,newObjLw,1);
}
bool samewb(const Param &tparam,const Param& param,double delta){
for(int i=0;i<tparam.w.size();i++){
if(fabs(tparam.w[i]-param.w[i])>delta){
return false;
}
}
if(fabs(tparam.d-param.d)>delta){
return false;
}
return true;
}
void displayIterProcess(int iter,double objLw,double newObjLw,int mod){
//每mod步打印一次迭代过程
if(iter%mod==0){
cout<<"迭代"<<iter<<":目标函数值【"<<newObjLw<<"】,两次迭代目标函数差值【 "<<(newObjLw-objLw)<<"】"<<endl;
cout<<"模型参数:";
for(int i=0;i<param.w.size();i++){
cout<<param.w[i]<<" ";
}
cout<<param.d<<endl<<endl;
}
}
//梯度上升更新w和b
void gradient(double lam){
for(int i=0;i<param.w.size();i++){
double tmp=0.0L;//保存梯度上升过程的中间值
for(int j=0;j<dataSet.size();j++){
tmp+=(dataSet[j].cls-logiFun(param,dataSet[j]))*dataSet[j].features[i]*lam;
}
param.w[i]+=(tmp);
}
double tmp=0.0L;
for(int j=0;j<dataSet.size();j++){
tmp+=(dataSet[j].cls-logiFun(param,dataSet[j]))*lam;
}
param.d+=tmp;
}
//计算logistic函数的值,即f(x)=exp(wx)/(1+exp(wx)),该表达式在求解梯度过程中出现,
//因此计算这个值是为了辅助梯度上升计算过程
inline double logiFun(const Param &p,const Data &d){
double inner=innerWX(p,d);
double le=exp(inner)/(1+exp(inner));
return le;
}
//计算对数似然函数的值
double Lw(Param p){
double l=0.0L;
for(int i=0;i<dataSet.size();i++){
double inner=innerWX(p,dataSet[i]);
l+=(dataSet[i].cls*inner-(log10(1+exp(inner))));
//cout<<"l="<<l<<endl;
}
return l;
}
//计算wx+b的值
inline double innerWX(const Param &p,const Data &data){
if(p.w.size()!=data.features.size()){
cout<<"参数与实例的维度不匹配,不能进行内积计算"<<endl;
exit(0);
}
double innerP=0.0L;
for(int i=0;i<p.w.size();i++){
innerP+=(p.w[i]*data.features[i]);
}
innerP+=p.d;
return innerP;
}
//给定测试集,预测分类
void predictClass(){
vector<Data> testDataSet;
loadDataSet(testDataSet,"./testdata.txt");
/*******************
分别计算
P(Y=1|x)=exp(w.x)/(1+exp(w.x))
和
P(Y=0|x)=1/(1+exp(w.x))
然后取值大的作为x的分类
*******************/
cout<<endl<<"预测分类:"<<endl;
for(int i=0;i<testDataSet.size();i++){
double py1=0.0L;
double py0=0.0L;
double inner=innerWX(param,testDataSet[i]);
py1=exp(inner)/(1+exp(inner));
py0=1-py1;
cout<<"实例: ";
for(int j=0;j<testDataSet[i].features.size();j++){
cout<<testDataSet[i].features[j]<<" ";
}
cout<<"标记分类【"<<testDataSet[i].cls<<"】,";
if(py1>=py0){
cout<<"预测分类【"<<1<<"】"<<endl;
}else{
cout<<"预测分类【"<<0<<"】"<<endl;
}
}
}
private:
vector<Data> dataSet;
Param param;
};
int main(){
Logistic logist;
//logist.displayDataSet();
logist.logisticRegression();
logist.predictClass();
system("pause");
return 0;
}
程序运行结果如下:
- 迭代训练模型过程如下
2. 测试分类过程如下:
本例中分类测试的效果并不是太好,在调试的时候,看到这样的结果我也仔细地审查代码,并未发现程序的错误(也希望读者帮忙审查)。考虑可能的原因:
(1)数据太少
(2)数据特征不能提供足够信息
(3)线性模型,无法很好地划分数据
附件:
训练数据:traindata.txt
10009 1 0 0 1 0 1
10025 0 0 1 2 0 0
10038 1 0 0 1 1 0
10042 0 0 0 0 1 0
10049 0 0 1 0 0 0
10113 0 0 1 0 1 0
10131 0 0 1 2 1 0
10160 1 0 0 0 0 0
10164 0 0 1 0 1 0
10189 1 0 1 0 0 0
10215 0 0 1 0 1 0
10216 0 0 1 0 0 0
10235 0 0 1 0 1 0
10270 1 0 0 1 0 0
10282 1 0 0 0 1 0
10303 2 0 0 0 1 0
10346 1 0 0 2 1 0
10380 2 0 0 0 1 0
10429 2 0 1 0 0 0
10441 0 0 1 0 1 0
10443 0 0 1 2 0 0
10463 0 0 0 0 0 0
10475 0 0 1 0 1 0
10489 1 0 1 0 1 1
10518 0 0 1 2 1 0
10529 1 0 1 0 0 0
10545 0 0 1 0 0 0
10546 0 0 0 2 0 0
10575 1 0 0 0 1 0
10579 2 0 1 0 0 0
10581 2 0 1 1 1 0
10600 1 0 1 1 0 0
10627 1 0 1 2 0 0
10653 1 0 0 1 1 0
10664 0 0 0 0 1 0
10691 1 1 0 0 1 0
10692 1 0 1 2 1 0
10711 0 0 0 0 1 0
10714 0 0 1 0 0 0
10739 1 0 1 1 1 0
10750 1 0 1 0 1 0
10764 2 0 1 2 0 0
10770 0 0 1 2 1 0
10780 0 0 1 0 1 0
10784 2 0 1 0 1 0
10785 0 0 1 0 1 0
10788 1 0 0 0 0 0
10815 1 0 0 0 1 0
10816 0 0 0 0 1 0
10818 0 0 1 2 1 0
11095 0 1 1 0 0 0
11146 0 1 0 0 1 0
11206 2 1 0 0 0 0
11223 2 1 0 0 0 0
11236 1 1 0 2 0 0
11244 1 1 0 0 0 1
11245 0 1 0 0 0 0
11278 2 1 0 0 1 0
11322 0 1 0 0 1 0
11326 2 1 0 2 1 0
11329 2 1 0 2 1 0
11344 1 1 0 2 1 0
11358 0 1 0 0 0 1
11417 2 1 1 0 1 0
11421 2 1 0 1 1 0
11484 1 1 0 0 0 1
11499 2 1 0 0 0 0
11503 1 1 0 0 1 0
11527 1 1 0 0 0 0
11540 2 1 0 1 1 0
11580 1 1 0 0 1 0
11583 1 0 1 1 0 1
11592 2 1 0 1 1 0
11604 0 1 0 0 1 0
11625 1 0 1 0 0 0
20035 0 0 1 0 0 1
20053 1 0 0 0 0 0
20070 0 0 0 2 1 0
20074 1 0 1 2 0 1
20146 1 0 0 1 1 0
20149 2 0 1 2 1 0
20158 2 0 0 0 1 0
20185 1 0 0 1 1 0
20193 1 0 1 0 1 0
20194 0 0 1 0 0 0
20205 1 0 0 2 1 0
20206 2 0 1 1 1 0
20265 0 0 1 0 1 0
20311 0 0 0 0 1 0
20328 2 0 0 1 0 1
20353 0 0 1 0 0 0
20372 0 0 0 0 0 0
20405 1 0 1 1 1 1
20413 2 0 1 0 1 0
20427 0 0 0 0 0 0
20455 1 0 1 0 1 0
20462 0 0 0 0 1 0
20472 0 0 0 2 0 0
20485 0 0 0 0 0 0
20523 0 0 1 2 0 0
20539 0 0 1 0 1 0
20554 0 0 1 0 0 1
20565 0 0 0 2 1 0
20566 1 0 1 1 1 0
20567 1 0 0 1 1 0
20568 0 0 1 0 1 0
20569 1 0 0 0 0 0
20571 1 0 1 0 1 0
20581 2 0 0 0 1 0
20583 1 0 0 0 1 0
20585 2 0 0 1 1 0
20586 0 0 1 2 1 0
20591 1 0 1 2 0 0
20595 0 0 1 2 1 0
20597 1 0 0 0 0 0
20599 0 0 1 0 1 0
20607 0 0 0 1 1 0
20611 1 0 0 0 1 0
20612 2 0 0 1 1 0
20614 1 0 0 1 1 0
20615 1 0 1 0 0 0
21017 1 1 0 1 1 0
21058 2 1 0 0 1 0
21063 0 1 0 0 0 0
21084 1 1 0 1 0 1
21087 1 1 0 2 1 0
21098 0 1 0 0 0 0
21099 1 1 0 2 0 0
21113 0 1 0 0 1 0
21114 1 1 0 0 1 1
21116 1 1 0 2 1 0
21117 1 0 0 2 1 0
21138 2 1 1 1 1 0
21154 0 1 0 0 1 0
21165 0 1 0 0 1 0
21181 2 1 0 0 0 1
21183 1 1 0 2 1 0
21231 1 1 0 0 1 0
21234 1 1 1 0 0 0
21286 2 1 0 2 1 0
21352 2 1 1 1 0 0
21395 0 1 0 0 1 0
21417 1 1 0 2 1 0
21423 0 1 0 0 1 0
21426 1 1 0 1 1 0
21433 0 1 0 0 1 0
21435 0 1 0 0 0 0
21436 1 1 0 0 0 0
21439 1 1 0 2 1 0
21446 1 1 0 0 0 0
21448 0 1 1 2 0 0
21453 2 1 0 0 1 0
30042 2 0 1 0 0 1
30080 0 0 1 0 1 0
301003 1 0 1 0 0 0
301009 0 0 1 2 1 0
301017 0 0 1 0 0 0
30154 1 0 1 0 1 0
30176 0 0 1 0 1 0
30210 0 0 1 0 1 0
30239 1 0 1 0 1 0
30311 0 0 0 0 0 1
30382 0 0 1 2 1 0
30387 0 0 1 0 1 0
30415 0 0 1 0 1 0
30428 0 0 1 0 0 0
30479 0 0 1 0 0 1
30485 0 0 1 2 1 0
30493 2 0 1 2 1 0
30519 0 0 1 0 1 0
30532 0 0 1 0 1 0
30541 0 0 1 0 1 0
30567 1 0 0 0 0 0
30569 2 0 1 1 1 0
30578 0 0 1 0 0 1
30579 1 0 1 0 0 0
30596 1 0 1 1 1 0
30597 1 0 1 1 0 0
30618 0 0 1 0 0 0
30622 1 0 1 1 1 0
30627 1 0 1 2 0 0
30648 2 0 0 0 1 0
30655 0 0 1 0 0 1
30658 0 0 1 0 1 0
30667 0 0 1 0 1 0
30678 1 0 1 0 0 0
30701 0 0 1 0 0 0
30703 2 0 1 1 0 0
30710 0 0 1 2 0 0
30713 1 0 0 1 1 1
30716 0 0 0 0 1 0
30721 0 0 0 0 0 1
30723 0 0 1 0 1 0
30724 2 0 1 2 1 0
30733 1 0 0 1 0 0
30734 0 0 1 0 0 0
30736 2 0 0 1 1 1
30737 0 0 1 0 0 0
30740 0 0 1 0 1 0
30742 2 0 1 0 1 0
30743 0 0 1 0 1 0
30745 2 0 0 0 1 0
30754 1 0 1 0 1 0
30758 1 0 0 0 1 0
30764 0 0 1 0 0 1
30765 2 0 0 0 0 0
30769 2 0 0 1 1 0
30772 0 0 1 0 1 0
30774 0 0 0 0 1 0
30784 2 0 1 0 0 0
30786 1 0 1 0 1 0
30787 0 0 0 0 1 0
30789 1 0 1 0 1 0
30800 0 0 1 0 0 0
30801 1 0 1 0 1 0
30803 1 0 1 0 1 0
30806 1 0 1 0 1 0
30817 0 0 1 2 0 0
30819 2 0 1 0 1 1
30822 0 0 1 0 1 0
30823 0 0 1 2 1 0
30834 0 0 0 0 0 0
30836 0 0 1 0 1 0
30837 1 0 1 0 1 0
30840 0 0 1 0 1 0
30841 1 0 1 0 0 0
30844 0 0 1 0 1 0
30845 0 0 1 0 0 0
30847 1 0 1 0 0 0
30848 0 0 1 0 1 0
30850 0 0 1 0 1 0
30856 1 0 0 0 1 0
30858 0 0 1 0 0 0
30860 0 0 0 0 1 0
30862 1 0 1 1 1 0
30864 0 0 0 2 0 0
30867 0 0 1 0 1 0
30869 0 0 1 0 1 0
30887 0 0 1 0 1 0
30900 1 0 0 1 1 0
30913 2 0 0 0 1 0
30914 1 0 0 0 0 0
30922 2 0 0 2 1 0
30923 0 0 1 2 1 0
30927 1 0 1 0 0 1
30929 0 0 1 2 1 0
30933 0 0 1 2 1 0
30940 0 0 1 0 1 0
30943 1 0 1 2 1 0
30945 0 0 0 2 0 0
30951 1 0 0 0 0 0
30964 0 0 0 2 1 0
30969 0 0 1 0 1 0
30979 2 0 0 0 1 0
30980 1 0 0 0 0 0
30982 1 0 0 1 1 0
30990 1 0 1 1 1 0
30991 1 0 1 0 1 1
30999 0 0 1 0 1 0
31056 1 1 0 2 1 0
31068 1 1 0 1 0 0
31108 2 1 0 2 1 0
31168 1 1 1 0 0 0
31191 0 1 1 0 0 0
31229 0 1 1 0 0 1
31263 0 1 0 0 1 0
31281 1 1 1 0 0 0
31340 1 1 1 0 1 0
31375 0 1 0 0 1 0
31401 0 1 1 0 0 1
31480 1 1 1 1 1 0
31501 1 1 0 2 1 0
31514 0 1 0 2 0 0
31518 1 1 0 2 1 0
31532 0 0 1 2 1 0
31543 2 1 1 1 1 0
31588 0 1 0 0 1 0
31590 0 0 1 0 1 0
31591 2 1 0 1 1 0
31595 0 1 0 0 1 0
31596 1 1 0 0 0 0
31598 1 1 0 0 1 0
31599 0 1 0 0 0 0
31605 0 1 1 0 0 0
31612 2 1 0 0 1 0
31615 2 1 0 0 0 0
31628 1 1 0 0 1 0
31640 2 1 0 1 1 0
测试数据:testdata.txt
10009 1 0 0 1 0 1
10025 0 0 1 2 0 0
20035 0 0 1 0 0 1
20053 1 0 0 0 0 0
30627 1 0 1 2 0 0
30648 2 0 0 0 1 0