15. 示例讲解
读懂符号系统, 剩下的就是技术了.
15.1 概率矩阵分解
预备知识: 高斯 (正态) 分布
概率密度函数
p
(
x
∣
μ
,
σ
)
=
1
2
π
σ
exp
(
−
(
x
−
μ
)
2
2
σ
2
)
(1)
p(x \vert \mu, \sigma) = \frac{1}{\sqrt{2 \pi} \sigma} \exp \left(- \frac{(x - \mu)^2}{2 \sigma^2}\right) \tag{1}
p(x∣μ,σ)=2πσ1exp(−2σ2(x−μ)2)(1)
概率矩阵分解论文: Ruslan Salakhutdinov and Andriy Mnih, Probabilistic Matrix Factorization.
p
(
R
∣
U
,
V
,
σ
)
=
∏
i
=
1
n
∏
j
=
1
m
[
N
(
r
i
j
∣
u
i
v
j
T
,
σ
)
]
I
i
j
(2)
p(\mathbf{R} \vert \mathbf{U}, \mathbf{V}, \sigma) = \prod_{i = 1}^n \prod_{j = 1}^m \left[ \mathcal{N}\left(r_{ij} \vert \mathbf{u}_i \mathbf{v}_j^\mathbf{T}, \sigma \right)\right]^{I_{ij}} \tag{2}
p(R∣U,V,σ)=i=1∏nj=1∏m[N(rij∣uivjT,σ)]Iij(2)
where
- R = [ r i j ] n × m \mathbf{R} = [r_{ij}]_{n \times m} R=[rij]n×m,
- U = [ u i j ] n × k \mathbf{U} = [u_{ij}]_{n \times k} U=[uij]n×k, V = [ v i j ] m × k \mathbf{V} = [v_{ij}]_{m \times k} V=[vij]m×k,
- I i j = 1 I_{ij} = 1 Iij=1 if r i j > 0 r_{ij} > 0 rij>0 and 0 0 0 otherwise.
15.2 多示例学习
Min-Ling Zhang and Zhi-Hua Zhou, Multi-Instance Clustering with Applications to Multi-Instance Prediction
Let
A
\mathbf{A}
A and
B
\mathbf{B}
B be two bags,
m
a
x
H
(
A
,
B
)
=
max
{
max
a
∈
A
min
b
∈
B
∥
a
−
b
∥
2
,
max
b
∈
B
min
a
∈
A
∥
b
−
a
∥
2
}
(5)
\mathrm{maxH}(\mathbf{A}, \mathbf{B}) = \max \{\max_{a \in \mathbf{A}} \min_{b \in \mathbf{B}} \|a - b\|_2, \max_{b \in \mathbf{B}} \min_{a \in \mathbf{A}} \|b - a\|_2\} \tag{5}
maxH(A,B)=max{a∈Amaxb∈Bmin∥a−b∥2,b∈Bmaxa∈Amin∥b−a∥2}(5)
m
i
n
H
(
A
,
B
)
=
min
a
∈
A
,
b
∈
B
∥
a
−
b
∥
2
(6)
\mathrm{minH}(\mathbf{A}, \mathbf{B}) = \min_{a \in \mathbf{A}, b \in \mathbf{B}} \|a - b\|_2 \tag{6}
minH(A,B)=a∈A,b∈Bmin∥a−b∥2(6)
- 原文没有下标 2, 但说了是欧氏距离.
- 分析的过程, 类似于展开多重括号.
15.3 作业
找一篇你们小组的论文来详细分析数学表达式, 包括其涵义, 规范, 优点和缺点.