偏置-方差 分解(bias-variance decomposition)
最近看论文需要借助的一个方法,这里简略地总结一下它地参数表。
原文:
P. Domingos, “A unified bias-variance decomposition and its applications,”
Proc. ICML, 2000, pp. 231–238
Training set | D = { ( x 1 , t 1 ) , . . , ( x n , t n ) } D=\{(x_1,t_1),..,(x_n,t_n)\} D={(x1,t1),..,(xn,tn)} |
---|---|
test example, model, prediction, true value | x x x, f f f, y = f ( x ) y=f(x) y=f(x), t t t |
loss function example | L ( t , y ) L(t,y) L(t,y) |
Optimal prediction | y ∗ y_* y∗ which minimizes E t [ L ( t , y ∗ ) ] E_t[L(t,y_*)] Et[L(t,y∗)]. |
main prediction | y m L , D = a r g m i n y ′ E D [ L ( y , y ′ ) ] y_m^{L,D}=argmin_{y'}E_D[L(y,y')] ymL,D=argminy′ED[L(y,y′)] |
bias | B ( x ) = L ( y ∗ , y m ) B(x)=L(y_*,y_m) B(x)=L(y∗,ym) |
variance | V ( x ) = E D [ L ( y m , y ) ] = E x [ B ( x ) ] V(x)=E_D[L(y_m,y)]=E_x[B(x)] V(x)=ED[L(ym,y)]=Ex[B(x)] |
noise | N ( x ) = E t [ L ( t , y ∗ ) ] N(x)=E_t[L(t,y_*)] N(x)=Et[L(t,y∗)] |
我们有:
期
望
误
差
=
方
差
+
偏
置
+
噪
声
E
D
,
t
[
L
(
t
,
y
)
]
=
c
1
E
t
[
L
(
t
,
y
∗
)
]
+
L
(
y
∗
,
y
m
)
+
c
2
E
D
[
L
(
y
m
,
y
)
]
=
c
1
N
(
x
)
+
B
(
x
)
+
c
2
V
(
x
)
\begin{aligned} 期望误差&=方差+偏置+噪声\\ E_{D,t}[L(t,y)]&=c_1E_t[L(t,y_*)]+L(y_*,y_m)+c_2E_D[L(y_m,y)]\\ &=c_1N(x)+B(x)+c_2V(x) \end{aligned}
期望误差ED,t[L(t,y)]=方差+偏置+噪声=c1Et[L(t,y∗)]+L(y∗,ym)+c2ED[L(ym,y)]=c1N(x)+B(x)+c2V(x)