引用Leo Breiman(1984) Classification_and_Regression_Trees 中数学论证:
Suppose a learning sample consisting of N cases (x 1 , y 1 ), ..., (x N , y N ) was used to construct a
predictor d(x)
Define the mean squared error R*(d) of the predictor d as
R*(d) = E(Y - d(X) )2 .
PROPOSITION8.2 . The predictor dB which minimizes R*(d) is
dB(x) = E(Y|X = x).
LEMMA 8.4. The constant a which minimizes
E(Y - a)2
is E(Y)
## 由上可知,回归,让MSE (或者平方根RMSE)最小,就可得最佳回归模型。。。
## 拟合优度 好象未见这样的最优数学基础。。
|