Random Walks, Random Fields, and Graph Kernels

Transcription

1 Random Walks, Random Fields, and Graph Kernels John Lafferty School of Computer Science Carnegie Mellon University Based on work with Avrim Blum, Zoubin Ghahramani, Risi Kondor Mugizi Rwebangira, Jerry Zhu

2 Outline Graph Kernels Random Fields Random Walks Continuous Fields

3 Using a Kernel ˆf(x) = N i= α i y i x, x i ˆf(x) = N i= α i y i K(x, x i ) 2

4 K(x, x ) positive semidefinite: The Kernel Trick X X f(x)f(x )K(x, x ) dx dx 0 Taking feature space of functions F = {Φ(x) = K(, x), x X }, has reproducing property g(x) = K(, x), g. Φ(x), Φ(x ) = K(, x), K(, x ) = K(x, x ) 3

5 Structured Data What if data lies on a graph or other data structure? S Google Cornell CMU N time VP foobar.com NSF L like V flies 4

6 Combinatorial Laplacian Think of edge e as tangent vector at e. For f : V R, df : E R is the -form df(e) = f(e + ) f(e ) Then = d d (as matrix) is discrete analogue of div 5

7 It is an averaging operator Combinatorial Laplacian f(x) = y x w xy (f(x) f(y)) = d(x) f(x) x y w xy f(y) We say f is harmonic if f = 0. Since f, g = df, dg, is self-adjoint and positive. 6

8 Diffusion Kernels on Graphs (Kondor and L., 2002) If is the graph Laplacian, in analogy with the continuous setting, t K t = K t is the heat equation on a graph. Solution K t = e t is the diffusion kernel. 7

9 Physical Interpretation ( t) K = 0, initial condition δx (y): e t f(x) = M K t(x, y) f(y) dy For a kernel-based classifier ŷ(x) = i α i y i K t (x i, x) decision function is given by heat flow with initial condition α i x = x i positive labeled data f(x) = α i x = x i negative labeled data 0 otherwise 8

10 RKHS Representation General spectral representation of a kernel as K(x, y) = n i= λ iφ i (x)φ i (y) leads to reproducing kernel Hilbert space i a i φ i, i b i φ i HK = i a i b i λ i For the diffusion kernel, RKHS inner product is f, g HK = i e tµ i fi ĝ i Interpretation: Functions with small norm don t oscillate rapidly on the graph. 9

11 Building Up Kernels If K (i) t are kernels on X i K t = n i= K(i) t is a kernel on X... X n. For the hypercube: Hamming distance {}}{ K t (x, x ) (tanh t) d(x, x ) Similar kernels apply to standard categorical data. Other graphs with explicit diffusion kernels: Infinite trees (Chung & Yau, 999) Rooted trees Cycles Strings with wildcards 0

12 Results on UCI Datasets Hamming Diffusion Kernel Improv. Data Set error SV error SV β err SV Breast Cancer 7.64% % % 83% Hepatitis 7.98% % % 58% Income 9.9% % % 8% Mushroom 3.36% % % 70% Votes 4.69% % % 2% Recent application to protein classification by Vert and Kanehisa (NIPS 2002).

13 Random Fields View of Combining Labeled/Unlabeled Data 2

14 Random Fields View View each vertex x as having label f(x) {+, }. Ising model on graph/lattice, spins f : V {+, } Energy H(f) = 2 w xy (f(x) f(y)) 2 x y x y w xy f(x) f(y) Gibbs distribution P (f) = Z(β) e βh(f) β = T Partition function Z(β) = f e βh(f) 3

15 Graph Mincuts Graph mincuts can be very unbalanced Graph mincuts don t exploit probabilistic properties of random fields Idea: Replace by averages under Ising model E β [f(x)] = f(x) e βh(f) Z(β) f S =f B 4

16 Pinned Ising Model β=3 β= β= β= β= β=

17 Not (Provably) Efficient to Approximate Unfortunately, analogue of rapid mixing result of Jerrum & Sinclair for ferromagnetic Ising model not known for mixed boundary conditions Question: Can we compute averages using graph algorithms in the zero temperature limit? 6

18 Idea: Relax to Statistical Field Theory Euclidean field theory on graph/lattice, fields f : V R Energy H(f) = 2 Gibbs distribution P (f) = Partition function Z(β) = x y w xy (f(x) f(y)) 2 Z(β) e βh(f) e βh(f) df f β = T Physical Interpretation: analytic continuation to imaginary time, t it Poincaré group Euclidean group. 7

19 View from Statistical Field Theory (cont.) Most probable field is harmonic Weighted graph G = (V, E), edge weights w xy, combinatorial Laplacian. Subgraph S with boundary S. Dirichlet Problem: unique solution f = 0 on S f S = f B 8

20 Random Walk Solution Perform random walk on unlabeled data, stop when hit a labeled point. What is the probability of hitting a positive labeled point before a negative labeled point? Precisely the same as minimum energy (continuous) random field. Label Propagation. Related work by Szummer and Jaakkola (NIPS 200) 9

21 Unconstrained Constrained

22 View from Statistical Field Theory In one-dimensional case: low temperature limit of average Ising model is the same is minimum energy Euclidean field. (Landau) Intuition: linear. average over graph s-t mincuts; harmonic solution is Not true in general... 2

23 Computing the Partition Function Let λ i be spectrum of, Dirichlet boundary conditions: Z(β) = e βh(f ) (βπ) n/2 det det = n i= λ i By generalization of matrix-tree (Chung & Langlands, 96) det = # rooted spanning forests i deg(i) 22

24 Connection with Diffusion Kernels Again take, combinatorial Laplacian with Dirichlet boundary conditions (zero on labeled data) For K t = e t diffusion kernel let K = 0 K t dt Solution to the Dirichlet problem (label prop, minimum energy continuous field): f (x) = z fringe K(x, z) f D (z) 23

25 Connection with Diffusion Kernels (cont.) Want to solve Laplace s equation: f = g. terms of. Solution given in Quick way to see connection using spectral representation: x,x = i µ i φ i (x) φ i (x ) K t (x, x ) = i e tµ i φ i (x) φ i (x ) x,x = i µ i φ i (x) φ i (x ) = 0 K t (x, x ) dt Used by Chung and Yau (2000). 24

26 Bounds on Covering Numbers and Generalization Error, Continuous Case Eigenvalue bounds from differential geometry (Li and Yau): )2 d µj c 2 ( j + )2 d c ( j V V Give bounds on SVM hypothesis class covering numbers log N (ɛ, F R (x)) = O (( V t d 2 ) log d+2 2 ( )) ɛ 25

27 Bounds on Generalization Error Better bounds on generalization error are now available based on Rademacher averages involving trace of the kernel (Bartlett, Bousquet, & Mendelson, preprint). Question: Can diffusion kernel connection be exploited to get transductive generalization error bounds for random walks approach? 26

28 Summary Random fields with discrete class labels intractable, unstable Continuous fields tractable, more desirable behavior for segentation and labeling Intimate connections with random walks, electric networks, graph flows, and diffusion kernels Advantages/disadvantages? 27