线性回归+多项式回归---中国人寿保费预测

数据准备与分析

读取

import numpy as np
import seaborn as sns
import pandas as pd
data = pd.read_excel("./中国人寿.xlsx")
data.head(5)

	age	sex	bmi	children	smoker	region	charges
0	19	female	27.900	0	yes	southwest	16884.92400
1	18	male	33.770	1	no	southeast	1725.55230
2	28	male	33.000	3	no	southeast	4449.46200
3	33	male	22.705	0	no	northwest	21984.47061
4	32	male	28.880	0	no	northwest	3866.85520

可以看到特征有6个：age、sex、bmi、children、smoker、region

分析

利用seaborn画出各个特征的概率分布曲线，删去不重要的特征

import matplotlib.pyplot as plt
fig, axes = plt.subplots(2, 2, figsize=(6, 4))
sns.kdeplot(data=data, x='charges', fill=True, hue='smoker',ax=axes[0,0])
sns.kdeplot(data=data, x='charges', fill=True, hue='children',ax=axes[0,1])
sns.kdeplot(data=data, x='charges', fill=True, hue='sex',ax=axes[1,0])
sns.kdeplot(data=data, x='charges', fill=True, hue='region',ax=axes[1,1])

分析:概率分布曲线重合度越高表明，特征的区别度越低

我们建模的时候可以把这些特征删去！

如上图，可以删去sex，和region这两个特征。

删去区别度低的特征

import warnings
warnings.filterwarnings("ignore")

val_data = data[['age','bmi','children','smoker','charges']]
## 字符数据编码
val_data['smoker'] = data['smoker'].apply(lambda x:0 if x == 'no' else 1)

X = val_data.iloc[:,:-1]
y = val_data[['charges']]
X

	age	bmi	children	smoker
0	19	27.900	0	1
1	18	33.770	1	0
2	28	33.000	3	0
3	33	22.705	0	0
4	32	28.880	0	0
...	...	...	...	...
1333	50	30.970	3	0
1334	18	31.920	0	0
1335	18	36.850	0	0
1336	21	25.800	0	0
1337	61	29.070	0	1

1338 rows × 4 columns

## 切分训练集与测试集

#定义一个预处理的类
class data_process:
    def __init__(self,X,y,train_partion=0.8):
        self.X = X
        self.y = y
        self.train_partion = train_partion
      
    #切分训练集与测试集
    def shuffle_data(self):
        #首先将数据转为np格式
        X = self.X.values
        y = self.y.values
        #打乱数据先
        index = np.arange(X.shape[0])
        np.random.shuffle(index)
        #numpy的花式索引，把样本打乱给成index的排序
        X = X[index]
        y = y[index]
        #按比例切分训练集和测试集
        margin = int(X.shape[0] * self.train_partion)
        X_train = X[:margin,:]
        y_train = y[:margin,:]
        X_test = X[margin:,:]
        y_test = y[margin:,:]
        return X_train,y_train,X_test,y_test
    #数据归一化
    def data_normolize(self):
        #0均值标准化
        #axis = 0列
        X_train,y_train,X_test,y_test = self.shuffle_data()
        X_train_norml = (X_train - X_train.mean(axis = 0)) / X_train.std(axis = 0)
        X_test_norml = (X_test - X_test.mean(axis = 0)) / X_test.std(axis = 0)
        return X_train_norml,y_train,X_test_norml,y_test
    def processing(self):
        X_train_norml,y_train,X_test_norml,y_test = self.data_normolize()
        return X_train_norml,y_train,X_test_norml,y_test

拿到处理完的数据

data_pro = data_process(X,y,train_partion=0.8)
X_train,y_train,X_test,y_test = data_pro.processing()
#display(X_train.shape,y_train.shape,X_test.shape,y_test.shape)

训练

使用线性回归、线性回归衍生算法
使用数据升维运算（对比不使用升维运算差异）

#R2分数评估模型，越接近1，模型越好
from sklearn.metrics import mean_squared_error,r2_score

线性回归

from sklearn.linear_model import LinearRegression

lnr = LinearRegression()
#训练
lnr.fit(X_train,y_train)
#预测
lnr_pre = lnr.predict(X_test)
#评分
lnr_score = r2_score(y_test,lnr_pre)
print("线性回归R2:",lnr_score)

线性回归R2: 0.7134052168884122

随机梯度下降

from sklearn.linear_model import SGDRegressor
sgd = SGDRegressor()
sgd.fit(X_train,y_train)
#预测
sgd_pre = sgd.predict(X_test)
#评分
sgd_score = r2_score(y_test,sgd_pre)
print("随机梯度下降R2:",sgd_score)

随机梯度下降R2: 0.7133771647173909

岭回归

from sklearn.linear_model import Ridge
ridge = Ridge(alpha= 1, solver='sag')
ridge.fit(X_train, y_train)
ridge_pre = ridge.predict(X_test)
#评分
ridge_score = r2_score(y_test,ridge_pre)
print("岭回归R2:",ridge_score)

岭回归R2: 0.7135328179378375

Lasso回归

from sklearn.linear_model import Lasso
lasso = Lasso(alpha= 0.5)
lasso.fit(X_train, y_train)
lasso_pre = lasso.predict(X_test)
lasso_score = r2_score(y_test,lasso_pre)
print("Lasso回归R2:",lasso_score)

Lasso回归R2: 0.7134145997065494

Elastic-Net

from sklearn.linear_model import ElasticNet
ElasticNet_model = ElasticNet(alpha= 1, l1_ratio = 0.7)
ElasticNet_model.fit(X_train, y_train)
ElasticNet_pre = ElasticNet_model.predict(X_test)
ElasticNet_score = r2_score(y_test,ElasticNet_pre)
print("ElasticNet R2:",ElasticNet_score)

ElasticNet R2: 0.698820757089488

数据升维

2次特征 ----- 是特征之间乘法+特征自身做升维运算的效果

# PolynomialFeatures，多项式，升维
from sklearn.preprocessing import PolynomialFeatures,StandardScaler
#没有归一化的数据
X_train,y_train,X_test,y_test = data_pro.shuffle_data()

poly = PolynomialFeatures(degree=2,interaction_only=False)#特征之间乘法+特征自身做升维运算的效果
X_train_poly = poly.fit_transform(X_train)
X_text_poly = poly.fit_transform(X_test)
#升完维再做归一化
stardard = StandardScaler() # 标准化
X_train_poly_norm = stardard.fit_transform(X_train_poly)
X_test_poly_norm = stardard.fit_transform(X_text_poly)

升维+线性回归

from sklearn.linear_model import LinearRegression
lnr_poly = LinearRegression()
#训练
lnr_poly.fit(X_train_poly_norm,y_train)
#预测
lnr_poly_pre = lnr_poly.predict(X_test_poly_norm)
#评分
lnr_score_ploy = r2_score(y_test,lnr_poly_pre)
print("二维多项式线性回归 R2:",lnr_score_ploy)

二维多项式线性回归 R2: 0.8743729177611113

升维+随机梯度下降

from sklearn.linear_model import SGDRegressor
sgd_ploy = SGDRegressor()
#训练
sgd_ploy.fit(X_train_poly_norm,y_train)
#预测
sgd_poly_pre = sgd_ploy.predict(X_test_poly_norm)
#评分
sgd_score_ploy = r2_score(y_test,sgd_poly_pre)
print("二维多项式随机梯度下降 R2:",sgd_score_ploy)

二维多项式随机梯度下降 R2: 0.8706325203066192

升维+岭回归

from sklearn.linear_model import Ridge
ridge_ploy = Ridge(alpha= 1, solver='sag')
ridge_ploy.fit(X_train_poly_norm, y_train)
ridge_pre_ploy = ridge_ploy.predict(X_test_poly_norm)
#评分
ridge_score_ploy = r2_score(y_test,ridge_pre_ploy)
print("升维+岭回归 R2:",ridge_score_ploy)

升维+岭回归 R2: 0.8743792671047281

升维+Lossa回归

from sklearn.linear_model import Lasso
lasso_ploy = Lasso(alpha= 0.5)
lasso_ploy.fit(X_train_poly_norm, y_train)
lasso_pre_ploy = lasso_ploy.predict(X_test_poly_norm)
#评分
lasso_score_ploy = r2_score(y_test,lasso_pre_ploy)
print("升维+Lossa回归 R2:",lasso_score_ploy)

升维+Lossa回归 R2: 0.8747696868608776

升维+Elastic-Net

from sklearn.linear_model import ElasticNet
ElasticNet_model_ploy = ElasticNet(alpha= 1, l1_ratio = 0.7)
ElasticNet_model_ploy.fit(X_train_poly_norm, y_train)
ElasticNet_pre_ploy = ElasticNet_model_ploy.predict(X_test_poly_norm)
#评分
ElasticNet_score_ploy = r2_score(y_test,ElasticNet_pre_ploy)
print("升维+Elastic-Net R2:",ElasticNet_score_ploy)

升维+Elastic-Net R2: 0.8104941823304773

总结:升维+岭回归的和升维+岭回归得分最高,这俩的分数差不多。试过三次的多项式得分不太理想，根据奥卡姆剃刀原则越简单越好，2次项回归针对本数据集可能是最好的。而且升维方式：是特征之间乘法+特征自身做升维运算的效果，比仅做特征之间的升维效果好！