Smooth Support Vector Regression - Python实现

算法特征:
①. 所有点尽可能落在管道内; ②. 极大化margin; ③. 极小化管道残差.
算法推导:
Part Ⅰ
引入光滑化手段: 详见https://www.cnblogs.com/xxhbdk/p/12275567.html
定义:
\begin{equation*}
\begin{split}
\left| x \right|_{\epsilon} &= \max\{0, \left| x \right| - \epsilon\} \\
&=\max\{0, x - \epsilon\} + \max\{0, -x - \epsilon\} \\
&=(x-\epsilon)_+ + (-x-\epsilon)_+
\end{split}
\end{equation*}
由于$(x-\epsilon)_+ \cdot (-x-\epsilon)_+ = 0$, 则
\begin{equation*}
\left|x\right|_{\epsilon}^2 = (x-\epsilon)_+^2 + (-x - \epsilon)_+^2
\end{equation*}
定义:
\begin{equation*}
p_{\epsilon}^2(x, \beta) = (p(x-\epsilon, \beta))^2 + (p(-x-\epsilon, \beta))^2
\end{equation*}
其中, $p(x, \beta) = \frac{1}{\beta}\ln(\mathrm{e}^{\beta x} + 1)$, $\beta$为光滑化参数.
现采用$p_{\epsilon}^2(x, \beta)$近似$\left|x\right|_{\epsilon}^2$, 相关函数图像如下:
下面给出$p_{\epsilon}^2(x, \beta)$的1阶及2阶导数:
\begin{align*}
\frac{\mathrm{d}p_{\epsilon}^2(x, \beta)}{\mathrm{d}x} &= 2[p(x-\epsilon, \beta)s(x-\epsilon, \beta) - p(-x-\epsilon, \beta)s(-x-\epsilon, \beta)] \\
\frac{\mathrm{d}^2p_{\epsilon}^2(x, \beta)}{\mathrm{d}x^2} &= 2[s^2(x-\epsilon, \beta) + p(x-\epsilon, \beta)\delta(x-\epsilon, \beta) + s^2(-x-\epsilon, \beta) + p(-x-\epsilon, \beta)\delta(-x-\epsilon, \beta)]
\end{align*}
其中, $s(x, \beta) = 1 / (\mathrm{e}^{-\beta x} + 1)$, $\delta(x, \beta) = \beta \mathrm{e}^{\beta x} / (\mathrm{e}^{\beta x} + 1)^2$.

Part Ⅱ
SVC线性回归方程如下:
\begin{equation}
h(x) = W^\mathrm{T}x + b
\label{eq_1}
\end{equation}
其中, $W$是超平面法向量, $b$是超平面偏置参数.
本文拟采用$L_2$范数衡量拟合残差, 则SVR原始优化问题如下:
\begin{equation}
\begin{split}
\min\quad &\frac{1}{2}\|W\|_2^2 + \frac{c}{2}(\|\xi\|_2^2 + \|\xi^{\ast}\|_2^2) \\
\mathrm{s.t.}\quad & \bar{Y} - (A^{\mathrm{T}}W + \mathbf{1}b) \leq \mathbf{1}\epsilon + \xi \\
& \bar{Y} - (A^{\mathrm{T}}W + \mathbf{1}b) \geq -\mathbf{1}\epsilon - \xi^{\ast}
\end{split}
\label{eq_2}
\end{equation}
其中, $A = [x^{(1)}, x^{(2)}, \cdots, x^{(n)}]$, $\bar{Y} = [\bar{y}^{(1)}, \bar{y}^{(2)}, \cdots, \bar{y}^{(n)}]^{\mathrm{T}}$, $\epsilon$是管道残差参数, $\xi$是由所有样本上管道残差组成的列矢量, $-\xi^{\ast}$是由所有样本下管道残差组成的列矢量, $c$为权重参数.
根据上述优化问题(\ref{eq_2}), 管道残差$\xi$、$\xi^{\ast}$可采用plus function等效表示:
\begin{equation}
\begin{cases}
\xi = (\bar{Y} - (A^{\mathrm{T}}W + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \\
\xi^{\ast} = (- \bar{Y} + (A^{\mathrm{T}}W + \mathbf{1}b) - \mathbf{1}\epsilon)_+
\end{cases}
\label{eq_3}
\end{equation}
由此可将该有约束最优化问题转化为如下无约束最优化问题:
\begin{equation}
\min\quad \frac{1}{2}\|W\|_2^2 + \frac{c}{2}(\| (\bar{Y} - (A^{\mathrm{T}}W + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2 + \| (- \bar{Y} + (A^{\mathrm{T}}W + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2)
\label{eq_4}
\end{equation}
同样, 根据优化问题(\ref{eq_2})KKT条件中关于原变量$W$的稳定性条件有:
\begin{equation}
W = A(\alpha - \alpha^{\ast}) = A\alpha^{\prime}
\label{eq_5}
\end{equation}
其中, $\alpha$、$\alpha^{\ast}$均为对偶变量. 代入式(\ref{eq_4})变数变换可得:
\begin{equation}
\min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}A^{\mathrm{T}}A\alpha^{\prime} + \frac{c}{2}(\| (\bar{Y} - (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2 + \| (- \bar{Y} + (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2)
\label{eq_6}
\end{equation}
由于权重参数$c$的存在, 上述最优化问题等效如下多目标最优化问题:
\begin{equation}
\begin{cases}
\min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}A^{\mathrm{T}}A\alpha^{\prime} \\
\min\quad \frac{1}{2}(\| (\bar{Y} - (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2 + \| (- \bar{Y} + (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2)
\end{cases}
\label{eq_7}
\end{equation}
由于$A^{\mathrm{T}}A\succeq 0$, 有如下等效:
\begin{equation}
\min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}A^{\mathrm{T}}A\alpha^{\prime} \quad\Rightarrow\quad \min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}\alpha^{\prime}
\label{eq_8}
\end{equation}
根据Part Ⅰ中光滑化手段, 有如下等效:
\begin{equation}
\min\quad \frac{1}{2}(\| (\bar{Y} - (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2 + \| (- \bar{Y} + (A^{\mathrm{T}}A\alpha^{\prime} + \mathbf{1}b) - \mathbf{1}\epsilon)_+ \|_2^2) \quad\Rightarrow\quad \min\quad \frac{1}{2}\sum_ip_{\epsilon}^2(\bar{y}^{(i)} - x^{(i)\mathrm{T}}A\alpha^{\prime} - b, \beta), \quad \text{where } \beta\rightarrow\infty
\label{eq_9}
\end{equation}
结合式(\ref{eq_8})、式(\ref{eq_9}), 优化问题(\ref{eq_6})等效如下:
\begin{equation}
\min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}\alpha^{\prime} + \frac{c}{2}\sum_ip_{\epsilon}^2(\bar{y}^{(i)} - x^{(i)\mathrm{T}}A\alpha^{\prime} - b, \beta), \quad \text{where } \beta\rightarrow\infty
\label{eq_10}
\end{equation}
为增强模型的回归能力, 引入kernel trick, 得最终优化问题:
\begin{equation}
\min\quad \frac{1}{2}\alpha^{\prime\mathrm{T}}\alpha^{\prime} + \frac{c}{2}\sum_ip_{\epsilon}^2(\bar{y}^{(i)} - K(x^{(i)\mathrm{T}}, A)\alpha^{\prime} - b, \beta), \quad \text{where } \beta\rightarrow\infty
\label{eq_11}
\end{equation}
其中, $K$代表kernel矩阵. 内部元素$K_{ij}=k(x^{(i)}, x^{(j)})$由kernel函数决定. 常见得kernel函数如下:
①. 多项式核: $k(x^{(i)}, x^{(j)}) = (x^{(i)}\cdot x^{(j)})^n$
②. 高斯核: $k(x^{(i)}, x^{(j)}) = \mathrm{e}^{-\mu\|x^{(i)} - x^{(j)}\|_2^2}$
结合式(\ref{eq_1})、式(\ref{eq_5})及kernel trick可得最终回归方程:
\begin{equation}
h(x) = \alpha^{\prime\mathrm{T}}K(A^{\mathrm{T}}, x) + b
\label{eq_12}
\end{equation}

Part Ⅲ
以下给出优化相关协助信息, 拟设式(\ref{eq_11})之目标函数为$J$, 并令:
\begin{equation}
J = J_1 + J_2 \quad\quad\text{其中,}\begin{cases}
J_1 = \frac{1}{2}\alpha^{\prime\mathrm{T}}\alpha^{\prime} \\
J_2 = \frac{c}{2}\sum\limits_ip_{\epsilon}^2(\bar{y}^{(i)} - K(x^{(i)\mathrm{T}}, A)\alpha^{\prime} - b, \beta)
\end{cases}
\label{eq_13}
\end{equation}
$J_1$相关:
\begin{gather*}
\frac{\partial J_1}{\partial \alpha^{\prime}} = \alpha^{\prime} \quad
\frac{\partial J_1}{\partial b} = 0 \\
\frac{\partial^2 J_1}{\partial \alpha^{\prime 2}} = I \quad
\frac{\partial^2 J_1}{\partial \alpha^{\prime} \partial b} = 0 \quad
\frac{\partial^2 J_1}{\partial b \partial \alpha^{\prime}} = 0 \quad
\frac{\partial^2 J_1}{\partial b^2} = 0
\end{gather*}
\begin{gather*}
\nabla J_1 = \begin{bmatrix}\frac{\partial J_1}{\partial \alpha^{\prime}} \\ \frac{\partial J_1}{\partial b} \end{bmatrix} \\ \\
\nabla^2 J_1 = \begin{bmatrix} \frac{\partial^2 J_1}{\partial \alpha^{\prime 2}} & \frac{\partial^2 J_1}{\partial \alpha^{\prime} \partial b} \\ \frac{\partial^2 J_1}{\partial b \partial \alpha^{\prime}} & \frac{\partial^2 J_1}{\partial b^2} \end{bmatrix}
\end{gather*}
$J_2$相关:
\begin{gather*}
z_i = \bar{y}^{(i)} - K(x^{(i)\mathrm{T}}, A)\alpha^{\prime} - b \\
\frac{\partial J_2}{\partial \alpha^{\prime}} = -c\sum_i[p(z_i-\epsilon, \beta)s(z_i-\epsilon, \beta) - p(-z_i - \epsilon, \beta)s(-z_i-\epsilon, \beta)]K^{\mathrm{T}}(x^{(i)\mathrm{T}}, A) \\
\frac{\partial J_2}{\partial b} = -c\sum_i[p(z_i-\epsilon, \beta)s(z_i-\epsilon, \beta) - p(-z_i - \epsilon, \beta)s(-z_i - \epsilon, \beta)] \\
\frac{\partial^2 J_2}{\partial \alpha^{\prime 2}} = c\sum_i[s^2(z_i-\epsilon, \beta) + p(z_i-\epsilon, \beta)\delta(z_i-\epsilon, \beta) + s^2(-z_i-\epsilon, \beta) + p(-z_i-\epsilon, \beta)\delta(-z_i-\epsilon, \beta)]K^{\mathrm{T}}(x^{(i)\mathrm{T}}, A)K(x^{(i)\mathrm{T}}, A) \\
\frac{\partial^2 J_2}{\partial \alpha^{\prime} \partial b} = c\sum_i[s^2(z_i-\epsilon, \beta) + p(z_i-\epsilon, \beta)\delta(z_i-\epsilon, \beta) + s^2(-z_i-\epsilon, \beta) + p(-z_i-\epsilon, \beta)\delta(-z_i-\epsilon, \beta)]K^{\mathrm{T}}(x^{(i)\mathrm{T}}, A) \\
\frac{\partial^2 J_2}{\partial b \partial \alpha^{\prime}} = c\sum_i[s^2(z_i-\epsilon, \beta) + p(z_i-\epsilon, \beta)\delta(z_i-\epsilon, \beta) + s^2(-z_i-\epsilon, \beta) + p(-z_i-\epsilon, \beta)\delta(-z_i-\epsilon, \beta)]K(x^{(i)\mathrm{T}}, A) \\
\frac{\partial^2 J_2}{\partial b^2} = c\sum_i[s^2(z_i-\epsilon, \beta) + p(z_i-\epsilon, \beta)\delta(z_i-\epsilon, \beta) + s^2(-z_i-\epsilon, \beta) + p(-z_i-\epsilon, \beta)\delta(-z_i-\epsilon, \beta)]
\end{gather*}
\begin{gather*}
\nabla J_2 = \begin{bmatrix}\frac{\partial J_2}{\partial \alpha^{\prime}} \\ \frac{\partial J_2}{\partial b} \end{bmatrix} \\ \\
\nabla^2 J_2 = \begin{bmatrix} \frac{\partial^2 J_2}{\partial \alpha^{\prime 2}} & \frac{\partial^2 J_2}{\partial \alpha^{\prime} \partial b} \\ \frac{\partial^2 J_2}{\partial b \partial \alpha^{\prime}}& \frac{\partial^2 J_2}{\partial b^2} \end{bmatrix}
\end{gather*}
目标函数$J$之gradient及Hessian如下:
\begin{align}
\nabla J &= \nabla J_1 + \nabla J_2 \\
\nabla^2 J &= \nabla^2 J_1 + \nabla^2 J_2
\end{align}

Part Ⅳ
现以如下peaks function为例进行算法实施:
\begin{equation*}
f(x_1, x_2) = 3(1 - x_1)^2\mathrm{e}^{-x_1^2 - (x_2 + 1)^2} - 10(\frac{x_1}{5} - x_1^3 - x_2^5)\mathrm{e}^{-x_1^2 - x_2^2} - \frac{1}{3}\mathrm{e}^{-(x_1 + 1)^2 - x_2^2}, \quad \text{where } x_1\in[-3, 3], x_2\in[-3, 3]
\end{equation*}
标准函数图像如下:

代码实现:

  1 # Smooth Support Vector Regression之实现
  2 
  3 import numpy
  4 from matplotlib import pyplot as plt
  5 from mpl_toolkits.mplot3d import Axes3D
  6 
  7 
  8 numpy.random.seed(0)
  9 
 10 def peaks(x1, x2):
 11     term1 = 3 * (1 - x1) ** 2 * numpy.exp(-x1 ** 2 - (x2 + 1) ** 2)
 12     term2 = -10 * (x1 / 5 - x1 ** 3 - x2 ** 5) * numpy.exp(-x1 ** 2 - x2 ** 2)
 13     term3 = -numpy.exp(-(x1 + 1) ** 2 - x2 ** 2) / 3
 14     val = term1 + term2 + term3
 15     return val
 16     
 17 
 18 # 生成回归数据
 19 X1 = numpy.linspace(-3, 3, 30)
 20 X2 = numpy.linspace(-3, 3, 30)
 21 X1, X2 = numpy.meshgrid(X1, X2)
 22 X = numpy.hstack((X1.reshape((-1, 1)), X2.reshape((-1, 1))))     # 待回归数据之样本点
 23 Y = peaks(X1, X2).reshape((-1, 1))
 24 Yerr = Y + numpy.random.normal(0, 0.4, size=Y.shape)             # 待回归数据之观测值
 25 
 26 
 27 
 28 class SSVR(object):
 29     
 30     def __init__(self, X, Y_, c=50, mu=1, epsilon=0.5, beta=10):
 31         '''
 32         X: 样本点数据集, 1行代表1个样本
 33         Y_: 观测值数据集, 1行代表1个观测值
 34         '''
 35         self.__X = X                                             # 待回归数据之样本点
 36         self.__Y_ = Y_                                           # 待回归数据之观测值
 37         self.__c = c                                             # 误差项权重参数
 38         self.__mu = mu                                           # gaussian kernel参数
 39         self.__epsilon = epsilon                                 # 管道残差参数
 40         self.__beta = beta                                       # 光滑化参数
 41         
 42         self.__A = X.T
 43         
 44     
 45     def get_estimation(self, x, alpha, b):
 46         '''
 47         获取估计值
 48         '''
 49         A = self.__A
 50         mu = self.__mu
 51         
 52         x = numpy.array(x).reshape((-1, 1))
 53         KAx = self.__get_KAx(A, x, mu)
 54         regVal = self.__calc_hVal(KAx, alpha, b)
 55         return regVal
 56         
 57         
 58     def get_MAE(self, alpha, b):
 59         '''
 60         获取平均绝对误差
 61         '''
 62         X = self.__X
 63         Y_ = self.__Y_
 64         
 65         Y = numpy.array(list(self.get_estimation(x, alpha, b) for x in X)).reshape((-1, 1))
 66         RES = Y_ - Y
 67         MAE = numpy.linalg.norm(RES, ord=1) / alpha.shape[0]
 68         return MAE
 69         
 70         
 71     def get_GE(self):
 72         '''
 73         获取泛化误差
 74         '''
 75         X = self.__X
 76         Y_ = self.__Y_
 77         
 78         cnt = 0
 79         GE = 0
 80         for idx in range(X.shape[0]):
 81             x = X[idx:idx+1, :]
 82             y_ = Y_[idx, 0]
 83             
 84             self.__X = numpy.vstack((X[:idx, :], X[idx+1:, :]))
 85             self.__Y_ = numpy.vstack((Y_[:idx, :], Y_[idx+1:, :]))
 86             self.__A = self.__X.T
 87             alpha, b, tab = self.optimize()
 88             if not tab:
 89                 continue
 90             cnt += 1
 91             y = self.get_estimation(x, alpha, b)
 92             GE += (y_ - y) ** 2
 93         GE /= cnt
 94         
 95         self.__X = X
 96         self.__Y_ = Y_
 97         self.__A = X.T
 98         return GE
 99         
100     
101     def optimize(self, maxIter=1000, EPSILON=1.e-9):
102         '''
103         maxIter: 最大迭代次数
104         EPSILON: 收敛判据, 梯度趋于0则收敛
105         '''
106         A, Y_ = self.__A, self.__Y_
107         c = self.__c
108         mu = self.__mu
109         epsilon = self.__epsilon
110         beta = self.__beta
111         
112         alpha, b = self.__init_alpha_b((A.shape[1], 1))
113         KAA = self.__get_KAA(A, mu)
114         JVal = self.__calc_JVal(KAA, Y_, c, epsilon, beta, alpha, b)
115         grad = self.__calc_grad(KAA, Y_, c, epsilon, beta, alpha, b)
116         Hess = self.__calc_Hess(KAA, Y_, c, epsilon, beta, alpha, b)
117         
118         for i in range(maxIter):
119             print("iterCnt: {:3d},   JVal: {}".format(i, JVal))
120             if self.__converged1(grad, EPSILON):
121                 return alpha, b, True
122                 
123             dCurr = -numpy.matmul(numpy.linalg.inv(Hess), grad)
124             ALPHA = self.__calc_ALPHA_by_ArmijoRule(alpha, b, JVal, grad, dCurr, KAA, Y_, c, epsilon, beta)
125             
126             delta = ALPHA * dCurr
127             alphaNew = alpha + delta[:-1, :]
128             bNew = b + delta[-1, -1]
129             JValNew = self.__calc_JVal(KAA, Y_, c, epsilon, beta, alphaNew, bNew)
130             if self.__converged2(delta, JValNew - JVal, EPSILON):
131                 return alphaNew, bNew, True
132             
133             alpha, b, JVal = alphaNew, bNew, JValNew
134             grad = self.__calc_grad(KAA, Y_, c, epsilon, beta, alpha, b)
135             Hess = self.__calc_Hess(KAA, Y_, c, epsilon, beta, alpha, b)
136         else:
137             if self.__converged1(grad, EPSILON):
138                 return alpha, b, True
139         return alpha, b, False
140             
141             
142     def __converged2(self, delta, JValDelta, EPSILON):
143         val1 = numpy.linalg.norm(delta)
144         val2 = numpy.linalg.norm(JValDelta)
145         if val1 <= EPSILON or val2 <= EPSILON:
146             return True
147         return False
148     
149     
150     def __calc_ALPHA_by_ArmijoRule(self, alphaCurr, bCurr, JCurr, gCurr, dCurr, KAA, Y_, c, epsilon, beta, C=1.e-4, v=0.5):
151         i = 0
152         ALPHA = v ** i
153         delta = ALPHA * dCurr
154         alphaNext = alphaCurr + delta[:-1, :]
155         bNext = bCurr + delta[-1, -1]
156         JNext = self.__calc_JVal(KAA, Y_, c, epsilon, beta, alphaNext, bNext)
157         while True:
158             if JNext <= JCurr + C * ALPHA * numpy.matmul(dCurr.T, gCurr)[0, 0]: break
159             i += 1
160             ALPHA = v ** i
161             delta = ALPHA * dCurr
162             alphaNext = alphaCurr + delta[:-1, :]
163             bNext = bCurr + delta[-1, -1]
164             JNext = self.__calc_JVal(KAA, Y_, c, epsilon, beta, alphaNext, bNext)
165         return ALPHA
166                 
167                 
168     def __converged1(self, grad, EPSILON):
169         if numpy.linalg.norm(grad) <= EPSILON:
170             return True
171         return False
172         
173         
174     def __calc_Hess(self, KAA, Y_, c, epsilon, beta, alpha, b):
175         Hess_J1 = self.__calc_Hess_J1(alpha)
176         Hess_J2 = self.__calc_Hess_J2(KAA, Y_, c, epsilon, beta, alpha, b)
177         Hess = Hess_J1 + Hess_J2
178         return Hess
179         
180         
181     def __calc_Hess_J2(self, KAA, Y_, c, epsilon, beta, alpha, b):
182         Hess_J2_alpha_alpha = numpy.zeros((KAA.shape[0], KAA.shape[0]))
183         Hess_J2_alpha_b = numpy.zeros((KAA.shape[0], 1))
184         Hess_J2_b_alpha = numpy.zeros((1, KAA.shape[0]))
185         Hess_J2_b_b = 0
186         
187         z = Y_ - numpy.matmul(KAA, alpha) - b
188         term1 = z - epsilon
189         term2 = -z - epsilon
190         for i in range(z.shape[0]):
191             term3 = self.__s(term1[i, 0], beta) ** 2 + self.__p(term1[i, 0], beta) * self.__d(term1[i, 0], beta)
192             term4 = self.__s(term2[i, 0], beta) ** 2 + self.__p(term2[i, 0], beta) * self.__d(term2[i, 0], beta)
193             term5 = term3 + term4
194             Hess_J2_alpha_alpha += term5 * numpy.matmul(KAA[:, i:i+1], KAA[i:i+1, :])
195             Hess_J2_alpha_b += term5 * KAA[:, i:i+1]
196             Hess_J2_b_b += term5
197         Hess_J2_alpha_alpha *= c
198         Hess_J2_alpha_b *= c
199         Hess_J2_b_alpha = Hess_J2_alpha_b.reshape((1, -1))
200         Hess_J2_b_b *= c
201 
202         Hess_J2_upper = numpy.hstack((Hess_J2_alpha_alpha, Hess_J2_alpha_b))
203         Hess_J2_lower = numpy.hstack((Hess_J2_b_alpha, [[Hess_J2_b_b]]))
204         Hess_J2 = numpy.vstack((Hess_J2_upper, Hess_J2_lower))
205         return Hess_J2
206         
207         
208     def __calc_Hess_J1(self, alpha):
209         I = numpy.identity(alpha.shape[0])
210         term = numpy.hstack((I, numpy.zeros((I.shape[0], 1))))
211         Hess_J1 = numpy.vstack((term, numpy.zeros((1, term.shape[1]))))
212         return Hess_J1
213         
214         
215     def __calc_grad(self, KAA, Y_, c, epsilon, beta, alpha, b):
216         grad_J1 = self.__calc_grad_J1(alpha)
217         grad_J2 = self.__calc_grad_J2(KAA, Y_, c, epsilon, beta, alpha, b)
218         grad = grad_J1 + grad_J2
219         return grad
220         
221         
222     def __calc_grad_J2(self, KAA, Y_, c, epsilon, beta, alpha, b):
223         grad_J2_alpha = numpy.zeros((KAA.shape[0], 1))
224         grad_J2_b = 0
225         
226         z = Y_ - numpy.matmul(KAA, alpha) - b
227         term1 = z - epsilon
228         term2 = -z - epsilon
229         for i in range(z.shape[0]):
230             term3 = self.__p(term1[i, 0], beta) * self.__s(term1[i, 0], beta) - self.__p(term2[i, 0], beta) * self.__s(term2[i, 0], beta)
231             grad_J2_alpha += term3 * KAA[:, i:i+1]
232             grad_J2_b += term3
233         grad_J2_alpha *= -c
234         grad_J2_b *= -c
235 
236         grad_J2 = numpy.vstack((grad_J2_alpha, [[grad_J2_b]]))
237         return grad_J2
238         
239         
240     def __calc_grad_J1(self, alpha):
241         grad_J1 = numpy.vstack((alpha, [[0]]))
242         return grad_J1
243         
244     
245     def __calc_JVal(self, KAA, Y_, c, epsilon, beta, alpha, b):
246         J1 = self.__calc_J1(alpha)
247         J2 = self.__calc_J2(KAA, Y_, c, epsilon, beta, alpha, b)
248         JVal = J1 + J2
249         return JVal
250         
251     
252     def __calc_J2(self, KAA, Y_, c, epsilon, beta, alpha, b):
253         z = Y_ - numpy.matmul(KAA, alpha) - b
254         term2 = numpy.sum(self.__p_epsilon_square(item[0], epsilon, beta) for item in z)
255         J2 = term2 * c / 2
256         return J2
257     
258         
259     def __calc_J1(self, alpha):
260         J1 = numpy.sum(alpha * alpha) / 2
261         return J1
262         
263         
264     def __p(self, x, beta):
265         val = numpy.log(numpy.exp(beta * x) + 1) / beta
266         return val
267         
268         
269     def __s(self, x, beta):
270         val = 1 / (numpy.exp(-beta * x) + 1)
271         return val
272         
273         
274     def __d(self, x, beta):
275         term = numpy.exp(beta * x)
276         val = beta * term / (term + 1) ** 2
277         return val
278         
279         
280     def __p_epsilon_square(self, x, epsilon, beta):
281         term1 = self.__p(x - epsilon, beta) ** 2
282         term2 = self.__p(-x - epsilon, beta) ** 2
283         val = term1 + term2
284         return val
285         
286     
287     def __get_KAA(self, A, mu):
288         KAA = numpy.zeros((A.shape[1], A.shape[1]))
289         for rowIdx in range(KAA.shape[0]):
290             for colIdx in range(rowIdx + 1):
291                 x1 = A[:, rowIdx:rowIdx+1]
292                 x2 = A[:, colIdx:colIdx+1]
293                 val = self.__calc_gaussian(x1, x2, mu)
294                 KAA[rowIdx, colIdx] = KAA[colIdx, rowIdx] = val
295         return KAA
296     
297     
298     def __calc_gaussian(self, x1, x2, mu):
299         val = numpy.exp(-mu * numpy.linalg.norm(x1 - x2) ** 2)
300         # val = numpy.sum(x1 * x2)
301         return val
302         
303         
304     def __init_alpha_b(self, shape):
305         '''
306         alpha、b之初始化
307         '''
308         alpha, b = numpy.zeros(shape), 0
309         return alpha, b
310         
311         
312     def __calc_hVal(self, KAx, alpha, b):
313         hVal = numpy.matmul(alpha.T, KAx)[0, 0] + b
314         return hVal
315     
316     
317     def __get_KAx(self, A, x, mu):
318         KAx = numpy.zeros((A.shape[1], 1))
319         for rowIdx in range(KAx.shape[0]):
320             x1 = A[:, rowIdx:rowIdx+1]
321             val = self.__calc_gaussian(x1, x, mu)
322             KAx[rowIdx, 0] = val
323         return KAx
324         
325         
326         
327 class PeaksPlot(object):
328     
329     def peaks_plot(self, X, Y_, ssvrObj, alpha, b):
330         surfX1 = numpy.linspace(numpy.min(X[:, 0]), numpy.max(X[:, 0]), 50)
331         surfX2 = numpy.linspace(numpy.min(X[:, 1]), numpy.max(X[:, 1]), 50)
332         surfX1, surfX2 = numpy.meshgrid(surfX1, surfX2)
333         surfY_ = peaks(surfX1, surfX2)
334         
335         surfY = numpy.zeros(surfX1.shape)
336         for rowIdx in range(surfY.shape[0]):
337             for colIdx in range(surfY.shape[1]):
338                 surfY[rowIdx, colIdx] = ssvrObj.get_estimation((surfX1[rowIdx, colIdx], surfX2[rowIdx, colIdx]), alpha, b)
339         
340         fig = plt.figure(figsize=(10, 3))
341         ax1 = plt.subplot(1, 2, 1, projection="3d")
342         ax2 = plt.subplot(1, 2, 2, projection="3d")
343         
344         norm = plt.Normalize(surfY_.min(), surfY_.max())
345         colors = plt.cm.rainbow(norm(surfY_))
346         surf = ax1.plot_surface(surfX1, surfX2, surfY_, facecolors=colors, shade=True, alpha=0.5)
347         surf.set_facecolor("white")
348         ax1.scatter3D(X[:, 0:1], X[:, 1:2], Y_, s=1, c="r")
349         ax1.set(xlabel="$x_1$", ylabel="$x_2$", zlabel="$f(x_1, x_2)$", xlim=(-3, 3), ylim=(-3, 3), zlim=(-8, 8))
350         ax1.set(title="Original Function")
351         ax1.view_init(0, 0)
352         
353         norm2 = plt.Normalize(surfY.min(), surfY.max())
354         colors2 = plt.cm.rainbow(norm2(surfY))
355         surf2 = ax2.plot_surface(surfX1, surfX2, surfY, facecolors=colors2, shade=True, alpha=0.5)
356         surf2.set_facecolor("white")
357         ax2.scatter3D(X[:, 0:1], X[:, 1:2], Y_, s=1, c="r")
358         ax2.set(xlabel="$x_1$", ylabel="$x_2$", zlabel="$f(x_1, x_2)$", xlim=(-3, 3), ylim=(-3, 3), zlim=(-8, 8))
359         ax2.set(title="Estimated Function")
360         ax2.view_init(0, 0)
361         
362         fig.tight_layout()
363         fig.savefig("peaks_plot.png", dpi=100)
364         
365         
366         
367 if __name__ == "__main__":
368     ssvrObj = SSVR(X, Yerr, c=50, mu=1, epsilon=0.5)
369     alpha, b, tab = ssvrObj.optimize()
370     
371     ppObj = PeaksPlot()
372     ppObj.peaks_plot(X, Yerr, ssvrObj, alpha, b)

View Code

结果展示:
左侧为样本点分布及原始函数图像, 右侧为样本点分布及估计函数图像.
使用建议:
①. 极大化margin可使管道投影加粗, 进一步使回归样本点尽可能落在管道投影内部;
②. SSVR本质上求解的仍然是原问题, 此时$\alpha^{\prime}$仅作为变数变换引入, 无需满足KKT条件中对偶可行性;
③. 数据集是否需要归一化, 应按实际情况予以考虑.
参考文档:
Yuh-Jye Lee, Wen-Feng Hsieh, and Chien-Ming Huang. "$\epsilon$-SSVR: A Smooth Support Vector Machine for $\epsilon$-Insensitive Regression", IEEE Transactions on Knowledge and Data Engineering, VoL. 17, No. 5, (2005), 678-685.

原文地址：https://www.cnblogs.com/xxhbdk/p/12425701.html