1. Formulations
- 假设一个页面上只有两种资源曝光,而总曝光量 s h o w T show_T showT一定,其中视频资源曝光 s h o w P show_P showP,图集资源曝光 s h o w I show_I showI。
- 假设视频和图集的点击率 c t r P ctr_P ctrP和 c t r I ctr_I ctrI分别为 s h o w P show_P showP和 s h o w I show_I showI的减函数且缓慢下降,i.e. c t r P = f ( s h o w P ) ctr_P=f(show_P) ctrP=f(showP), c t r I = g ( s h o w I ) ctr_I=g(show_I) ctrI=g(showI),且 f ′ ( s h o w P ) < 0 f'(show_P)<0 f′(showP)<0, g ′ ( s h o w I ) < 0 g'(show_I)<0 g′(showI)<0, f ′ ′ ( s h o w P ) > 0 f''(show_P)>0 f′′(showP)>0, g ′ ′ ( s h o w I ) > 0 g''(show_I)>0 g′′(showI)>0;
- 假设视频和图集的次均时长 d u r P dur_P durP和 d u r I dur_I durI不随曝光数而改变;
- 整体的收益使用总时长计算,公式为:
Π = c t r P d u r P s h o w P + c t r I d u r I s h o w I = f ( s h o w P ) d u r P s h o w P + g ( s h o w I ) d u r I s h o w I = f ( s h o w P ) d u r P s h o w P + g ( s h o w T − s h o w P ) d u r I ( s h o w T − s h o w P ) \begin{aligned} \Pi&=ctr_Pdur_Pshow_P+ctr_Idur_Ishow_I\\ &=f(show_P)dur_Pshow_P+g(show_I)dur_Ishow_I\\ &=f(show_P)dur_Pshow_P+g(show_T-show_P)dur_I(show_T-show_P) \end{aligned} Π=ctrPdurPshowP+ctrIdurIshowI=f(showP)durPshowP+g(showI)durIshowI=f(showP)durPshowP+g(showT−showP)durI(showT−showP)
其中 s h o w T show_T showT 是一个常数。
2. 计算
D ( Π ) / D ( s h o w P ) = f ( s h o w P ) d u r P + d u r P s h o w P f ′ ( s h o w P ) − g ( s h o w T − s h o w P ) d u r I − d u r I ( s h o w T − s h o w P ) g ′ ( s h o w T − s h o w P ) \begin{aligned} D(\Pi)/D(show_P)&=f(show_P)dur_P+dur_Pshow_Pf'(show_P)\\ &-g(show_T-show_P)dur_I-dur_I(show_T-show_P)g'(show_T-show_P)\\ \end{aligned} D(Π)/D(showP)=f(showP)durP+durPshowPf′(showP)−g(showT−showP)durI−durI(showT−showP)g′(showT−showP)
I
1
=
f
(
s
h
o
w
P
)
+
s
h
o
w
P
f
′
(
s
h
o
w
P
)
I_1=f(show_P)+show_Pf'(show_P)
I1=f(showP)+showPf′(showP)
I
2
=
−
g
(
s
h
o
w
T
−
s
h
o
w
P
)
−
(
s
h
o
w
T
−
s
h
o
w
P
)
g
′
(
s
h
o
w
T
−
s
h
o
w
P
)
I_2=-g(show_T-show_P)-(show_T-show_P)g'(show_T-show_P)
I2=−g(showT−showP)−(showT−showP)g′(showT−showP)
D
(
I
1
)
/
D
(
s
h
o
w
P
)
=
2
f
′
(
s
h
o
w
P
)
+
s
h
o
w
P
f
′
′
(
s
h
o
w
P
)
D(I_1)/D(show_P)=2f'(show_P)+show_Pf''(show_P)
D(I1)/D(showP)=2f′(showP)+showPf′′(showP)
D
(
I
2
)
/
D
(
s
h
o
w
P
)
=
2
g
′
(
s
h
o
w
T
−
s
h
o
w
P
)
+
(
s
h
o
w
T
−
s
h
o
w
P
)
g
′
′
(
s
h
o
w
T
−
s
h
o
w
P
)
D(I_2)/D(show_P)=2g'(show_T-show_P)+(show_T-show_P)g''(show_T-show_P)
D(I2)/D(showP)=2g′(showT−showP)+(showT−showP)g′′(showT−showP)
发现在不清楚 f ′ ( s h o w P ) f'(show_P) f′(showP)与 f ′ ′ ( s h o w P ) f''(show_P) f′′(showP) 关系的情况下,上面两式的符号无法判断。
3. simulation
假设
f
(
s
h
o
w
P
)
=
a
−
b
∗
s
h
o
w
P
f(show_P)=a-b*show_P
f(showP)=a−b∗showP,
g
(
s
h
o
w
I
)
=
c
−
d
∗
s
h
o
w
I
g(show_I)=c-d*show_I
g(showI)=c−d∗showI
则:
D
(
Π
)
/
D
(
s
h
o
w
P
)
=
(
a
−
2
b
∗
s
h
o
w
P
)
d
u
r
P
−
(
c
−
2
d
∗
s
h
o
w
T
+
2
d
∗
s
h
o
w
P
)
d
u
r
I
D(\Pi)/D(show_P)=(a-2b*show_P)dur_P-(c-2d*show_T+2d*show_P)dur_I
D(Π)/D(showP)=(a−2b∗showP)durP−(c−2d∗showT+2d∗showP)durI
二阶导数 D 2 ( Π ) / D ( s h o w P ) 2 = − 2 b ∗ d u r P − 2 d ∗ d u r I < 0 D^2(\Pi)/D(show_P)^2=-2b*dur_P-2d*dur_I<0 D2(Π)/D(showP)2=−2b∗durP−2d∗durI<0
一阶导数的极限值:
-
s h o w P → 0 ⇒ D ( Π ) / D ( s h o w P ) → a d u r P − c ∗ d u r I + 2 d ∗ s h o w T ∗ d u r I ) > 0 show_P\rightarrow0\\\Rightarrow D(\Pi)/D(show_P)\rightarrow adur_P-c*dur_I+2d*show_T*dur_I)>0 showP→0⇒D(Π)/D(showP)→adurP−c∗durI+2d∗showT∗durI)>0
-
s h o w P → s h o w T ⇒ D ( Π ) / D ( s h o w P ) → ( a − 2 b ∗ s h o w T ) d u r P − c ∗ d u r I < 0 show_P\rightarrow show_T\\\Rightarrow D(\Pi)/D(show_P)\rightarrow (a-2b*show_T)dur_P-c*dur_I<0 showP→showT⇒D(Π)/D(showP)→(a−2b∗showT)durP−c∗durI<0
说明 Π \Pi Π 随着 s h o w P show_P showP 的增加,先涨后降,并且在 D ( Π ) / D ( s h o w P ) = 0 D(\Pi)/D(show_P)=0 D(Π)/D(showP)=0即 ( a − 2 b ∗ s h o w P ) d u r P − ( c − 2 d ∗ s h o w T + 2 d ∗ s h o w P ) d u r I = 0 (a-2b*show_P)dur_P-(c-2d*show_T+2d*show_P)dur_I=0 (a−2b∗showP)durP−(c−2d∗showT+2d∗showP)durI=0时取得极大值。
化简得:
s h o w P ∗ = a ∗ d u r P − c ∗ d u r I + 2 d ∗ s h o w T d u r I 2 b ∗ d u r P + 2 d ∗ d u r I = a ∗ d u r P d u r I − c + 2 d ∗ s h o w T 2 b ∗ d u r P d u r I + 2 d \begin{aligned} show_P^*&=\frac{a*dur_P-c*dur_I+2d*show_Tdur_I}{2b*dur_P+2d*dur_I}\\ &=\frac{a*\frac{dur_P}{dur_I}-c+2d*show_T}{2b*\frac{dur_P}{dur_I}+2d} \end{aligned} showP∗=2b∗durP+2d∗durIa∗durP−c∗durI+2d∗showTdurI=2b∗durIdurP+2da∗durIdurP−c+2d∗showT
这个结果的含义是:
- 视频 c t r ctr ctr 上限 a a a 越高,视频 c t r ctr ctr 随曝光衰减比率 b b b 越低,越倾向于曝光更多视频;
- 图集 c t r ctr ctr 上限 c c c 越高,图集 c t r ctr ctr 随曝光衰减比率 d d d 越低,越倾向于曝光更少视频;
- 二类资源次均时长差异 d u r P d u r I \frac{dur_P}{dur_I} durIdurP越高,越倾向于曝光更多视频。