简单线性回归:目标:找到a和b,使得 ∑ i = 1 m ( y t r a i n ( i ) − a x t r a i n ( i ) − b ) 2 \sum_{i=1}^m(y_{train}^{(i)}-ax_{train}^{(i)}-b)^2 ∑i=1m(ytrain(i)−axtrain(i)−b)2尽可能小
线性回归算法的评测:
衡量标准: ∑ i = 1 m ( y t e s t ( i ) − y ^ t e s t ( i ) ) 2 \sum_{i=1}^m(y_{test}^{(i)}-\hat{y}_{test}^{(i)})^2 ∑i=1m(ytest(i)−y^test(i))2,问题:和m相关?
1)均方误差MSE(Mean Squared Error)问题:量纲?
1
m
∑
i
=
1
m
(
y
t
e
s
t
(
i
)
−
y
^
t
e
s
t
(
i
)
)
2
\frac1m\sum_{i=1}^m(y_{test}^{(i)}-\hat{y}_{test}^{(i)})^2
m1i=1∑m(ytest(i)−y^test(i))2
2)均方根误差RMSE(Root Mean Squared Error)
1
m
∑
i
=
1
m
(
y
t
e
s
t
(
i
)
−
y
^
t
e
s
t
(
i
)
)
2
=
M
S
E
t
e
s
t
\sqrt{\frac1m\sum_{i=1}^m(y_{test}^{(i)}-\hat{y}_{test}^{(i)})^2}=\sqrt{MSE_{test}}
m1i=1∑m(ytest(i)−y^test(i))2=MSEtest
3)平方绝对误差MAE(Mean Absolute Error)
1
m
∑
i
=
1
m
∣
y
t
e
s
t
(
i
)
−
y
^
t
e
s
t
(
i
)
∣
\frac1m\sum_{i=1}^m\mid y_{test}^{(i)}-\hat{y}_{test}^{(i)}\mid
m1i=1∑m∣ytest(i)−y^test(i)∣
问题:分类的准确度:1最好,0最差
4)R Squared
R
2
=
1
−
S
S
r
e
s
i
d
u
a
l
S
S
t
o
t
a
l
R^2=1-\frac{SS_{residual}}{SS_{total}}
R2=1−SStotalSSresidual
Residual Sum of Squares;Total Sum of Squares
R
2
=
1
−
∑
i
(
y
^
(
i
)
−
y
(
i
)
)
2
∑
i
(
y
ˉ
−
y
(
i
)
)
2
R^2=1-\frac{\sum_i(\hat{y}^{(i)}-y^{(i)})^2}{\sum_i(\bar{y}-y^{(i)})^2}
R2=1−∑i(yˉ−y(i))2∑i(y^(i)−y(i))2
分子:使用我们的模型预测产生的错误;
分母:使用
y
=
y
ˉ
预
测
产
生
的
错
误
y=\bar{y}预测产生的错误
y=yˉ预测产生的错误
Baseline Model
1.
R
2
≤
1
R^2\leq 1
R2≤1
2.
R
2
R^2
R2越大越好。当我们的预测模型不犯任何错误时,
R
2
R^2
R2得到最大值1;
3.当我们的模型等于基准模型时,
R
2
R^2
R2为0;
4.如果
R
2
<
0
R^2<0
R2<0,说明我们学习到的模型还不如基准模型。此事,很有可能我们的数据不存在任何线性关系。
R
2
=
1
−
∑
i
(
y
^
(
i
)
−
y
(
i
)
)
2
∑
i
(
y
ˉ
−
y
(
i
)
)
2
R^2=1-\frac{\sum_i(\hat{y}^{(i)}-y^{(i)})^2}{\sum_i(\bar{y}-y^{(i)})^2}
R2=1−∑i(yˉ−y(i))2∑i(y^(i)−y(i))2
=
1
−
(
∑
i
m
(
y
^
(
i
)
−
y
(
i
)
)
2
)
/
m
(
∑
i
m
(
y
ˉ
−
y
(
i
)
)
2
)
/
m
=1-\frac{(\sum_i^m(\hat{y}^{(i)}-y^{(i)})^2)/m}{(\sum_i^m(\bar{y}-y^{(i)})^2)/m}
=1−(∑im(yˉ−y(i))2)/m(∑im(y^(i)−y(i))2)/m
=
1
−
M
S
E
(
y
^
,
y
)
V
a
r
(
y
)
=1-\frac{MSE(\hat{y},y)}{Var(y)}
=1−Var(y)MSE(y^,y)