总结
MLP做了一个广告主的采纳率模型(阿里20年才做广告主侧的优化吗。。。)
细节
用context bandit算法,学习advertisers’ preference以及recommendation adoption
ad campaign
v
=
[
v
1
,
v
2
,
…
,
v
n
]
T
\bold{v} = [v_1, v_2, \dots, v_n]^T
v=[v1,v2,…,vn]T,其中n
代表KPI的个数。
w
∗
\bold{w}^*
w∗代表广告主的偏好,目标函数:
∏
w
∗
=
arg max
∏
w
∗
T
⋅
v
∏
\prod_{\bold{w}^*} = \argmax_{\prod}\bold{w}^{*T} \cdot \bold{v}_{\prod}
w∗∏=∏argmaxw∗T⋅v∏
context bandit建模
status
S
\mathcal{S}
S:campaign相关的特征
action
A
\mathcal{A}
A:预测广告主的偏好
w
∗
{\bold{w}^*}
w∗
reward
R
\mathcal{R}
R:采纳为1,其他为0
cb的目标函数
R
(
T
)
=
E
[
∑
t
=
1
T
r
t
,
w
t
∗
]
−
E
[
∑
t
=
1
T
r
t
,
w
t
]
R(T) = \bold{E}[\sum_{t=1}^T r_{t, \bold{w}_t^*}] - \bold{E}[\sum_{t=1}^T r_{t, \bold{w}_t}]
R(T)=E[t=1∑Trt,wt∗]−E[t=1∑Trt,wt]
其中
w
t
∗
\bold{w}_t^*
wt∗表示最优偏好
本文的action是高维、连续的,用cb无法有效选择action,因此:
w
=
M
L
P
(
advertiser behaviours, ad profile, scenario, ...
)
\bold{w} = MLP(\text{advertiser behaviours, ad profile, scenario, ...})
w=MLP(advertiser behaviours, ad profile, scenario, ...)
loss function:
L
=
−
1
N
∑
(
x
,
v
,
y
)
∈
D
(
y
×
log
p
(
x
,
v
)
+
(
1
−
y
)
×
log
(
1
−
p
(
x
,
v
)
)
L = -\frac{1}{N} \sum_{(x, v, y) \in \mathcal{D}} (y \times \log p(x, v) + (1 - y) \times \log(1 - p(x, v))
L=−N1(x,v,y)∈D∑(y×logp(x,v)+(1−y)×log(1−p(x,v))
其中
v
v
v是ad performance vector,
x
x
x是environment feature,
y
y
y是label(是否采纳)
实验
实验组+对照组客户消费占全量30%,50/50区分的实验对照组
评估指标
在线实验
采纳率:实验组采纳率4.5%,对照组采纳率3.8%
ARPU/average revenue per user:实验组arpu+1.2%
离线实验
accumulated expected regret,也即上面的
R
(
T
)
R(T)
R(T)
accumulated adoption rate,
∑
t
=
1
T
r
t
,
w
t
∗
T
\frac{\sum_{t=1}^T r_{t, \bold{w}_t^*}}{T}
T∑t=1Trt,wt∗