基函数为回归数:
T(x;Θ)=j=1JcjI(xRj) T(x;\Theta) = \sum_{j=1}^Jc_jI(x \in R_j)

前向分步算法为:
fo(x)=0fm(x)=fm1(x)+T(x;Θm),m=1,2,,MfM(x)=m=1MT(x;Θm) \begin{aligned} f_o(x) = 0 \\ f_m(x) = f_{m-1}(x) + T(x;\Theta_m), m = 1,2,\cdots,M \\ f_M(x) = \sum_{m=1}^MT(x;\Theta_m) \end{aligned}

假设当前模型为fm1(x)f_{m-1}(x),解得第m棵树的参数为:
Θ^margminΘmi=1NL(yi,fm1(xi)+T(x;Θm)) \hat \Theta_m \arg\min_{\Theta_m}\sum_{i=1}^NL(y_i, f_{m-1}(x_i)+T(x;\Theta_m))

采用平方误差损失函数:
L(y,f(x))=(yf(x))2 L(y, f(x)) = (y-f(x))^2

其损失变为:
L(y,fm1(x)+T(x;Θm))=(yfm1(x)T(x;Θm))2 L(y, f_{m-1}(x)+T(x;\Theta_m)) = (y-f_{m-1}(x)-T(x;\Theta_m))^2
yfm1(x)=T(x;Θm)y-f_{m-1}(x)=T(x;\Theta_m)时,L最小。
yfm1(x)y-f_{m-1}(x)称为当前拟合数据的残差(residual)
直观解释为:新的回归树只是简单地拟合当前模型的残差。

results matching ""

    No results matching ""