- 逻辑回归模型参数详解
求解对偶问题：如果数据量小但特征很多可以改成true
一般可以设置max_iter 大一些而tol小一些
Class_weight:输入{0:1,1:3}则代表1类样本的每条数据在计算损失函数时都会*3，当输入balanced，则调整为真实样本比例的反比，以达到平衡，但实际情况中不常用
Multi_class:默认情况auto，模型会优先根据惩罚项和solver选择OVR还是MVM
Solver：
- 逻辑回归调参实验
首先介绍PolynomialFeatures函数，其参数有degree（最高阶数）、interaction_only（是否包含交叉项）、include_bias（是否只包含0阶计算结果、偏置项）
具体实验代码如下：
#!/usr/bin/env python
coding: utf-8
In[1]:
import numpy as np
import pandas as pd
In[2]:
import matplotlib as mpl
import matplotlib.pyplot as plt
In[3]:
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import PolynomialFeatures
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
In[4]:
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
In[5]:
np.random.seed(24)
x=np.random.normal(0,1,size=(1000,2))
In[6]:
y=np.array(x[:,0]+x[:,1]**2<1.5,int)
In[7]:
plt.scatter(x[:,0],x[:,1],c=y)#c表示颜色
In[8]:
np.random.seed(24)
for i in range(200):
y[np.random.randint(1000)]=1
y[np.random.randint(1000)]=0
In[9]:
plt.scatter(x[:,0],x[:,1],c=y)
In[10]:
x_train,x_test,y_train,y_test=train_test_split(x,y,train_size=0.7,random_state=42)
In[26]:
def plr(degree=1,penalty='none',C=1.0):
pipe=make_pipeline(PolynomialFeatures(degree=degree,include_bias=False),
StandardScaler(),
LogisticRegression(penalty=penalty,tol=1e-4,C=C,max_iter=int(1e4)))
return pipe
#UI多迭代10的6次方次，tol是优化算法的收敛容忍度，c是正则化项参数
In[27]:
pl1=plr()
In[28]:
pl1.fit(x_train,y_train)
In[29]:
pl1.score(x_train,y_train),pl1.score(x_test,y_test)
In[37]:
def plot_decision_boundary(x,y,model):
x1,x2=np.meshgrid(
np.linspace(x[:,0].min()-1,x[:,0].max()+1,1000).reshape(-1,1),
np.linspace(x[:,1].min()-1,x[:,1].max()+1,1000).reshape(-1,1))
x_temp=np.concatenate([x1.reshape(-1,1),x2.reshape(-1,1)],1)
yhat_temp=model.predict(x_temp)
yhat=yhat_temp.reshape(x1.shape)
from matplotlib.colors import ListedColormap
custom_cmap=ListedColormap(['#EF9A9A','#90CAF9'])
plt.contourf(x1,x2,yhat,cmap=custom_cmap)
plt.scatter(x[(y==0).flatten(),0],x[(y==0).flatten(),1],color='red')
plt.scatter(x[(y==1).flatten(),0],x[(y==1).flatten(),1],color='red')
In[38]:
plot_decision_boundary(x,y,pl1)
In[39]:
#再看下2次特征进行建模：
pr2=plr(degree=2)
In[40]:
pr2.fit(x_train,y_train)
In[41]:
pr2.score(x_train,y_train),pr2.score(x_test,y_test)
#(0.7914285714285715, 0.7866666666666666)，分数提升10%
In[43]:
plot_decision_boundary(x,y,pr2)
In[45]:
#如何查看参数情况
pr2.named_steps['logisticregression'].coef_
#array([[-0.81012988, 0.04384694, -0.48583038, 0.02977868, -1.12352417]])
In[46]:
#过拟合倾向实验
pr3=plr(degree=10)
pr3.fit(x_train,y_train)
pr3.score(x_train,y_train),pr3.score(x_test,y_test)
#(0.8314285714285714, 0.78)
In[47]:
plot_decision_boundary(x,y,pr3)
In[48]:
#尝试不同参数下准确率评分
score_l=[]
for degree in range(1,21):
pr_temp=plr(degree=degree)
pr_temp.fit(x_train,y_train)
score_temp=[pr_temp.score(x_train,y_train),pr_temp.score(x_test,y_test)]
score_l.append(score_temp)
In[49]:
np.array(score_l)
In[54]:
#画图看准确率变化
plt.plot(list(range(1,21)),np.array(score_l)[:,0],label='train_acc')
plt.plot(list(range(1,21)),np.array(score_l)[:,1],label='test_acc')
plt.legend(loc=4)#指定图例位置
In[55]:
#手动调参，尝试LL1正则化
pl1=plr(degree=10,penalty='l1',C=1.0)
In[57]:
pl1.set_params(logisticregression__solver='saga')
pl1.fit(x_train,y_train)#直接fit会报错，要改变求解器为saga
In[58]:
pl1.score(x_train,y_train),pl1.score(x_test,y_test)
In[62]:
#尝试枚举搜索参数有degree、C、正则化项
score_l1=[]
for degree in range(1,21):
pr_temp=plr(degree=degree,penalty='l1')
pr_temp.set_params(logisticregression__solver='saga')
pr_temp.fit(x_train,y_train)
score_temp=[pr_temp.score(x_train,y_train),pr_temp.score(x_test,y_test)]
score_l1.append(score_temp)
plt.plot(list(range(1,21)),np.array(score_l1)[:,0],label='train_acc')
plt.plot(list(range(1,21)),np.array(score_l1)[:,1],label='test_acc')
plt.legend(loc=4)#指定图例位置
In[63]:
score_l1#打印发现degree=3是最优解，以此为degree进行后面的搜索
In[64]:
score_l2=[]
for degree in range(1,21):
pr_temp=plr(degree=degree,penalty='l2')
pr_temp.set_params(logisticregression__solver='saga')
pr_temp.fit(x_train,y_train)
score_temp=[pr_temp.score(x_train,y_train),pr_temp.score(x_test,y_test)]
score_l2.append(score_temp)
plt.plot(list(range(1,21)),np.array(score_l2)[:,0],label='train_acc')
plt.plot(list(range(1,21)),np.array(score_l2)[:,1],label='test_acc')
plt.legend(loc=4)#指定图例位置
In[66]:
score_l2#打印发现degree=15是最优解，以此为degree进行后面的搜索
In[72]:
#尝试C的取值
score_l1_3=[]
for c in np.arange(0.5,2,0.1):
pr_temp=plr(degree=3,penalty='l1',C=c)
pr_temp.set_params(logisticregression__solver='saga')
pr_temp.fit(x_train,y_train)
score_temp=[pr_temp.score(x_train,y_train),pr_temp.score(x_test,y_test)]
score_l1_3.append(score_temp)
plt.plot(list(np.arange(0.5,2,0.1)),np.array(score_l1_3)[:,0],label='train_acc')
plt.plot(list(np.arange(0.5,2,0.1)),np.array(score_l1_3)[:,1],label='test_acc')
plt.legend(loc=4)#指定图例位置
In[73]:
score_l1_3#因此准确率最高为0.8左右