该论文注释21中，提及并简化了论文《Topic and role discovery in social networks with experiments on Enron and academic email》- A. McCallum, X. Wang & A. Corrada-Emmanuel.的式子：
P
(
x
d
i
,
z
d
i
∣
x
−
d
i
,
z
−
d
i
,
w
,
α
,
β
,
a
,
r
)
=
P
(
x
d
i
,
z
d
i
,
w
d
i
∣
x
−
d
i
,
z
−
d
i
,
w
−
d
i
,
α
,
β
,
a
,
r
)
P
(
w
d
i
∣
x
−
d
i
,
z
−
d
i
,
w
−
d
i
,
α
,
β
,
a
,
r
)
∝
P
(
x
,
z
,
w
∣
α
,
β
,
a
,
r
)
P
(
x
−
d
i
,
z
−
d
i
,
w
−
d
i
∣
α
,
β
,
a
,
r
)
∝
Γ
(
α
z
d
i
+
n
a
d
x
d
i
z
d
i
)
Γ
(
α
z
d
i
+
n
a
d
x
d
i
z
d
i
−
1
)
Γ
(
β
w
d
i
+
m
z
d
i
w
d
i
)
Γ
(
β
w
d
i
+
m
z
d
i
w
d
i
−
1
)
Γ
(
∑
t
=
1
T
(
α
t
+
n
a
d
x
d
i
t
)
)
Γ
(
∑
t
=
1
T
(
α
t
+
n
a
d
x
d
i
t
)
−
1
)
Γ
(
∑
v
=
1
V
(
β
v
+
m
z
d
i
v
)
)
Γ
(
∑
v
=
1
V
(
β
v
+
m
z
d
i
v
)
−
1
)
∝
α
z
d
i
+
n
a
d
x
d
i
z
d
i
−
1
∑
t
=
1
T
(
α
t
+
n
a
d
x
d
i
)
−
1
β
w
d
i
+
m
z
d
i
w
d
i
−
1
∑
v
=
1
(
β
v
+
m
z
d
i
v
)
−
1
\begin{alignedat}{3} &P\left(x_{d i}, z_{d i} \mid \mathbf{x}_{-d i}, \mathbf{z}_{-d i}, \mathbf{w}, \alpha, \beta, \mathbf{a}, \mathbf{r}\right) \\ &= \frac{P\left(x_{d i}, z_{d i}, w_{d i} \mid \mathbf{x}_{-d i}, \mathbf{z}_{-d i}, \mathbf{w}_{-d i}, \alpha, \beta, \mathbf{a}, \mathbf{r}\right)}{P\left(w_{d i} \mid \mathbf{x}_{-d i}, \mathbf{z}_{-d i}, \mathbf{w}_{-d i}, \alpha, \beta, \mathbf{a}, \mathbf{r}\right)} \propto \frac{P(\mathbf{x}, \mathbf{z}, \mathbf{w} \mid \alpha, \beta, \mathbf{a}, \mathbf{r})}{P\left(\mathbf{x}_{-d i}, \mathbf{z}_{-d i}, \mathbf{w}_{-d i} \mid \alpha, \beta, \mathbf{a}, \mathbf{r}\right)} \\ &\propto\frac{\frac{\Gamma\left(\alpha_{z_{d i}}+n_{a_{d} x_{d i} z_{d i}}\right)}{\Gamma\left(\alpha_{z_{d i}}+n_{\left.a_{d} x_{d i} z_{d i}-1\right)}\right.} \frac{\Gamma\left(\beta_{w_{d i}}+m_{z_{d i} w_{d i}}\right)}{\Gamma\left(\beta_{w_{d i}}+m_{\left.z_{d i} w_{d i}-1\right)}\right.} }{ \frac{\Gamma\left(\sum_{t=1}^{T}\left(\alpha_{t}+n_{a_{d} x_{d i} t}\right)\right)}{\Gamma\left(\sum_{t=1}^{T}\left(\alpha_{t}+n_{a_{d} x_{d i} t}\right)-1\right)} \frac{\Gamma\left(\sum_{v=1}^{V}\left(\beta_{v}+m_{z_{d i} v}\right)\right)}{\Gamma\left(\sum_{v=1}^{V}\left(\beta_{v}+m_{z_{d i} v}\right)-1\right)} }\propto\frac{\alpha_{z_{d i}}+n_{a_{d} x_{d i} z_{d i}}-1}{\sum_{t=1}^{T}\left(\alpha_{t}+n_{a_{d} x_{d i}}\right)-1} \frac{\beta_{w_{d i}}+m_{z_{d i} w_{d i}}-1}{\sum_{v=1}^{}\left(\beta_{v}+m_{z_{d i} v}\right)-1} \end{alignedat}
P(xdi,zdi∣x−di,z−di,w,α,β,a,r)=P(wdi∣x−di,z−di,w−di,α,β,a,r)P(xdi,zdi,wdi∣x−di,z−di,w−di,α,β,a,r)∝P(x−di,z−di,w−di∣α,β,a,r)P(x,z,w∣α,β,a,r)∝Γ(∑t=1T(αt+nadxdit)−1)Γ(∑t=1T(αt+nadxdit))Γ(∑v=1V(βv+mzdiv)−1)Γ(∑v=1V(βv+mzdiv))Γ(αzdi+nadxdizdi−1)Γ(αzdi+nadxdizdi)Γ(βwdi+mzdiwdi−1)Γ(βwdi+mzdiwdi)∝∑t=1T(αt+nadxdi)−1αzdi+nadxdizdi−1∑v=1(βv+mzdiv)−1βwdi+mzdiwdi−1
具体是将其简化成:
p
(
z
i
∣
z
⃗
¬
i
,
w
⃗
¬
i
)
=
p
(
z
i
,
w
i
∣
z
⃗
¬
i
,
w
⃗
¬
i
)
/
p
(
w
i
∣
z
⃗
¬
i
,
w
⃗
¬
i
)
∝
p
(
z
⃗
,
w
⃗
)
/
p
(
z
⃗
¬
i
,
w
⃗
¬
i
)
p(z_{i} \mid \vec{z}_{\neg i}, \vec{w}_{\neg i})=p(z_{i}, w_{i} \mid \vec{z}_{\neg i}, \vec{w}_{\neg i}) / p(w_{i} \mid \vec{z}_{\neg i}, \vec{w}_{\neg i}) \propto p(\vec{z}, \vec{w}) / p(\vec{z}_{\neg i}, \vec{w}_{\neg i})
p(zi∣z¬i,w¬i)=p(zi,wi∣z¬i,w¬i)/p(wi∣z¬i,w¬i)∝p(z,w)/p(z¬i,w¬i)
但这个简化写法似乎有误，应为：
p
(
z
i
∣
z
⃗
¬
i
,
w
⃗
)
=
p
(
z
i
,
w
i
∣
z
⃗
¬
i
,
w
⃗
¬
i
)
/
p
(
w
i
∣
z
⃗
¬
i
,
w
⃗
¬
i
)
∝
p
(
z
⃗
,
w
⃗
)
/
p
(
z
⃗
¬
i
,
w
⃗
¬
i
)
p(z_{i} \mid \vec{z}_{\neg i}, \vec{w})=p(z_{i}, w_{i} \mid \vec{z}_{\neg i}, \vec{w}_{\neg i}) / p(w_{i} \mid \vec{z}_{\neg i}, \vec{w}_{\neg i}) \propto p(\vec{z}, \vec{w}) / p(\vec{z}_{\neg i}, \vec{w}_{\neg i})
p(zi∣z¬i,w)=p(zi,wi∣z¬i,w¬i)/p(wi∣z¬i,w¬i)∝p(z,w)/p(z¬i,w¬i)
以及，本人对相关公式的推演草稿（未做整理）暂存如下：
扫一扫
3080
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
