Discrete regularity for graph Laplacians --Jeff Calder
pdf和视频资源:http://www.ipam.ucla.edu/abstract/?tid=15794&pcode=HJWS4
Introduction
- graph-based learning. 目的基本都是做clustering或者半监督的clustering。
geometric weights w x y = η ( ∣ x − y ∣ ε ) w_{xy}=\eta\left(\dfrac{|x-y|}{\varepsilon}\right) wxy=η(ε∣x−y∣) - spectral clustering.
– 最简单的 graph cuts: min A ⊂ χ Cut ( A ) : = ∑ x , y ∈ χ ; x ∈ A , y ∉ A w x y \min_{A \subset \chi} \text{Cut}(A): = \sum_{x,y\in \chi; x\in A, y \notin A} w_{xy} minA⊂χCut(A):=∑x,y∈χ;x∈A,y∈/Awxy. 问题:使 A = x A={x} A=x -> unbalanced
– balanced cut: min A ⊂ χ Cut ( A ) Vol ( A ) Vol ( χ ∖ A ) \min_{A \subset \chi} \dfrac{\text{Cut}(A)}{\text{Vol}(A)\text{Vol}(\chi \setminus A)} minA⊂χVol(A)Vol(χ∖A)Cut(A),其中 Vol ( A ) = ∑ x ∈ A ∑ y ∈ χ w x y \text{Vol}(A)=\sum_{x \in A}\sum_{y \in \chi} w_{xy} Vol(A)=∑x∈A∑y∈χwxy。问题:NP难
– spectral clustering:
设 u ( x ) = 1 u(x)=1 u(x)=1,若 x ∈ A x \in A x∈A,否则为0。有
Cut ( A ) = ∑ x , y ∈ χ ; x ∈ A , y ∉ A w x y = 1 2 ∑ x , y ∈ χ w x y ( u ( x ) − u ( y ) ) 2 \text{Cut}(A)=\sum_{x,y \in \chi; x \in A, y \notin A}w_{xy}=\dfrac{1}{2}\sum_{x,y \in \chi} w_{xy} (u(x)-u(y))^2 Cut(A)=x,y∈χ;x∈A,y∈/A∑wxy=21x,y∈χ∑wxy(u(x)−u(y))2 且 Vol ( A ) = ∑ x , y ∈ χ w x y u ( x ) \text{Vol}(A)=\sum_{x,y \in \chi} w_{xy} u(x) Vol(A)=∑x,y∈χwxyu(x)。此时,balanced cut可变为
min u : χ → { 0 , 1 } ∑ x , y ∈ χ w x y ( u ( x ) − u ( y ) ) 2 ∑ x , y , x ′ , y ′ ∈ χ u ( x ) w x y ( 1 − u ( y ′ ) ) w x ′ y ′ \min_{u:\chi \rightarrow \{0,1\}} \dfrac{\sum_{x,y \in \chi} w_{xy}(u(x)-u(y))^2}{\sum_{x,y,x',y' \in \chi} u(x)w_{xy} (1-u(y'))w_{x'y'}} u:χ→{0,1}min∑x,y,x′,y′∈χu(x)wxy(1−u(y′))wx′y′∑x,y∈χwxy(u(x)−u(y))2 min u : χ → R , ∑ x ∈ χ u ( x ) ≠ 0 ∑ x , y ∈ χ w x y ( u ( x ) − u ( y ) ) 2 ∑ x ∈ χ u ( x ) 2 \min_{u:\chi \rightarrow R, \sum_{x\in \chi} u(x) \neq 0} \dfrac{\sum_{x,y \in \chi} w_{xy}(u(x)-u(y))^2}{\sum_{x \in \chi} u(x)^2} u:χ→R,∑x∈χu(x)=0min∑x∈χu(x)2∑x,y∈χwxy(u(x)−u(y))2 第二个是第一个的relaxed problem,其解为拉普拉斯图的最小非平凡特征向量(Fiedler向量)。
分成 k k k 组聚类,去拉普拉斯前 k k k 个特征向量: u 1 , . . . , u k : χ → R u_1,...,u_k: \chi \rightarrow \mathbb{R} u1,...,uk:χ→R。定义spectral embedding Ψ : χ → R k \Psi:\chi\rightarrow \mathbb{R}^k Ψ:χ→Rk,有 Ψ ( x ) = ( u 1 ( x ) , . . . , u k ( x ) ) \Psi(x)=(u_1(x),...,u_k(x)) Ψ(x)=(u1(x),...,uk(x))。 - the manifold assumption
– 先简单了解下流形,见链接https://www.zhihu.com/question/24015486
– 设 M ⊂ R d \mathcal{M} \subset \mathbb{R}^d M⊂Rd 为compact, connected, orientable, smooth的 m m m维流形。 d M ( x , y ) d_{\mathcal{M}}(x,y) dM(x,y):the geodesic distance between x , y ∈ M x,y \in \mathcal{M} x,y∈M。定义一个球 B M ( x , r ) = { y ∈ M : d M ( x , y ) < r } B_{\mathcal{M}}(x,r)=\{y \in \mathcal{M}: d_{\mathcal{M}}(x,y)<r\} BM(x,r)={y∈M:dM(x,y)<r}.
密度 ρ ∈ C 2 ( M ) \rho\in C^2(\mathcal{M}) ρ∈C2(M), ρ > 0 \rho>0 ρ>0。令 χ n = { x 1 , . . . , x n } \chi_n=\{x_1,...,x_n\} χn={x1,...,xn}为 ρ d Vol M \rho d\textit{Vol}_{\mathcal{M}} ρdVolM的一个独立同分布。
– 随机几何图 random geometric graph,权重为 w x y = η ( ∣ x − y ∣ ε ) w_{xy}=\eta \left( \dfrac{|x-y|}{\varepsilon }\right) wxy=η(ε∣x−y∣),其中 η ( t ) = 0 , for t > 1 , ∫ R m η ( ∣ w ∣ ) d w = 1 \eta(t) = 0 \textit{, for } t>1, \int_{\mathbb{R}^m}\eta(|w|)dw=1 η(t)=0, for t>1,∫Rmη(∣w∣)dw=1 第二个条件为Lipschitz。
– spectral convergence. 图-拉普拉斯算子的谱收敛( n → ∞ , ε → 0 n \rightarrow \infty, \varepsilon \rightarrow 0 n→∞,ε→0)到加权Laplace-Beltrami算子的谱,即 Δ M u = − ρ − 1 div M ( ρ 2 ▽ M u ) \Delta_{\mathcal{M}}u=-\rho^{-1}\text{div}_{\mathcal{M}}(\rho^2\bigtriangledown_{\mathcal{M}}u) ΔMu=−ρ−1divM(ρ2▽Mu)