多项式回归

有 a和b 两个特征，那么它的 2 次多项式的次数为 1, a, b,  , ab,  

PolynomialFeatures 这个类有 3 个参数

degree：控制多项式的次数；

interaction_only：默认为 False，如果指定为 True，那么就不会有特征⾃⼰和⾃⼰结合的项，组合的特征中没有 a 和 b ；

include_bias：默认为 True 。如果为 True 的话，那么结果中就会有 0 次幂项，即全为 1 这⼀列。

from sklearn.preprocessing import PolynomialFeatures
import numpy as np
from sympy import false

X = np.arange(6).reshape(3, 2)
print(X)
poly = PolynomialFeatures(degree=2)
res = poly.fit_transform(X)
print(res)

结果

[[0 1] 
 [2 3] 
 [4 5]]

[[ 1.  0.  1.  0.  0.  1.] 
 [ 1.  2.  3.  4.  6.  9.] 
 [ 1.  4.  5. 16. 20. 25.]]

 

from sklearn.preprocessing import PolynomialFeatures
import numpy as np
from sympy import false
from sklearn.linear_model import LinearRegression
import matplotlib.pyplot as plt
from sklearn.metrics import mean_squared_error



data = np.array([[ -2.95507616, 10.94533252],
[ -0.44226119, 2.96705822],
[ -2.13294087, 6.57336839],
[ 1.84990823, 5.44244467],
[ 0.35139795, 2.83533936],
[ -1.77443098, 5.6800407 ],
[ -1.8657203 , 6.34470814],
[ 1.61526823, 4.77833358],
[ -2.38043687, 8.51887713],
[ -1.40513866, 4.18262786]])
X = data[:, 0].reshape(-1, 1) # 将array转换成矩阵 reshape(-1,1)转换成1列
y = data[:,1].reshape(-1, 1)

poly_features = PolynomialFeatures(degree=2, include_bias=False)
X_poly = poly_features.fit_transform(X)
print(X_poly)

lin_reg = LinearRegression()
lin_reg.fit(X_poly, y)
print(lin_reg.intercept_, lin_reg.coef_)
# [ 2.60996757] [[-0.12759678 0.9144504 ]]
X_plot = np.linspace(-3, 3, 1000).reshape(-1, 1)
X_plot_poly = poly_features.fit_transform(X_plot)
y_plot = lin_reg.predict(X_plot_poly)
plt.plot(X_plot, y_plot, 'red')
plt.plot(X, y, 'b.')
plt.show()

y_pre = lin_reg.predict(X_poly)
mean_squared_error(y, y_pre)

 

 结果：[2.60996757] [[-0.12759678  0.9144504 ]]

方程为为h = −0.13x + 0.91+ 2.61

 引⼊⾼阶项，训练误差明显下降，那么训练误差是否还有进⼀步下降的空间呢？

from sklearn.preprocessing import PolynomialFeatures
import numpy as np
from sympy import false
from sklearn.linear_model import LinearRegression
import matplotlib.pyplot as plt
from sklearn.metrics import mean_squared_error

data = np.array([[ -2.95507616, 10.94533252],
[ -0.44226119, 2.96705822],
[ -2.13294087, 6.57336839],
[ 1.84990823, 5.44244467],
[ 0.35139795, 2.83533936],
[ -1.77443098, 5.6800407 ],
[ -1.8657203 , 6.34470814],
[ 1.61526823, 4.77833358],
[ -2.38043687, 8.51887713],
[ -1.40513866, 4.18262786]])
X = data[:, 0].reshape(-1, 1) # 将array转换成矩阵 reshape(-1,1)转换成1列
y = data[:,1].reshape(-1, 1)
def try_degree(degree, X, y):
    poly_features_d = PolynomialFeatures(degree=degree, include_bias=False)
    X_poly_d = poly_features_d.fit_transform(X)
    lin_reg_d = LinearRegression()
    lin_reg_d.fit(X_poly_d, y)
    return {'X_poly': X_poly_d, 'intercept': lin_reg_d.intercept_, 'coef': lin_reg_d.coef_}
degree2loss_paras = []
for i in range(2, 20):
    paras = try_degree(i, X, y)
# ⾃⼰实现预测值的求解
h = np.dot(paras['X_poly'], paras['coef'].T) + paras['intercept']
_loss = mean_squared_error(h, y)
degree2loss_paras.append({'d': i, 'loss': _loss, 'coef': paras['coef'], 'intercept': paras['intercept']})
min_index = np.argmin(np.array([i['loss'] for i in degree2loss_paras]))
min_loss_para = degree2loss_paras[min_index]
print(min_loss_para)

X_plot = np.linspace(-3, 1.9, 1000).reshape(-1, 1)
poly_features_d = PolynomialFeatures(degree=min_loss_para['d'], include_bias=False)
X_plot_poly = poly_features_d.fit_transform(X_plot)
y_plot = np.dot(X_plot_poly, min_loss_para['coef'].T) + min_loss_para['intercept']
plt.plot(X_plot, y_plot, 'red', label="degree12")
plt.plot(X, y, 'b.', label="X")
plt.legend(loc='best')
plt.show()

 此时出现过拟合现象，鉴于该问题的普遍性和重要性，在满⾜要求的情况下，能选择简单模型时应该尽量选择简单的模型。