Lecture 9: Laplacian Eigenmaps

Transcription

1 Lecture 9: Radu Balan Department of Mathematics, AMSC, CSCAMM and NWC University of Maryland, College Park, MD April 18, 2017

2 Optimization Criteria Assume G = (V, W ) is a undirected weighted graph with n nodes and weight matrix W. We interpret W i,j as the similarity between nodes i and j. The larger the weight the more similar the nodes, and the closer they are in a geometric graph embedding. Thus we look for a dimension d > 0 and a set of points {y 1, y 2,, y n } R d so that d i,j = y i y j s is small for large weight W i,j.

3 Optimization Criteria Assume G = (V, W ) is a undirected weighted graph with n nodes and weight matrix W. We interpret W i,j as the similarity between nodes i and j. The larger the weight the more similar the nodes, and the closer they are in a geometric graph embedding. Thus we look for a dimension d > 0 and a set of points {y 1, y 2,, y n } R d so that d i,j = y i y j s is small for large weight W i,j. A natural optimization criterion candidate: J(y 1, y 2,, y n ) = 1 i,j n W i,j y i y j 2,

4 Optimization Criteria Lemma J(y 1, y 2,, y n ) = 1 2 is convex in (y 1,, y n ). W i,j y i y j 2 1 i,j n Proof Idea: Write it as a positive semidefinite quadratic criterion: n J = y i 2 n n W i,j W i,j y i, y j i=1 j=1 i,j=1 Let Y = [y 1 y n ] be the d n matrix of coordinates. Let D = diag(d k ), with d k = n i=1 W k,i, be a n n diagonal matrix. A little algebra shows: { } J(Y ) = trace Y (D W )Y T.

5 Optimization Criteria Equivalent forms: J(Y ) = trace { } { } Y (D W )Y T = trace Y Y T = trace { G} where G = Y T Y is the n n Gram matrix. Thus: J is quadratic in Y, and positive semidefnite, hence convex. Also: J is linear in G.

6 Optimization Criteria Equivalent forms: J(Y ) = trace { } { } Y (D W )Y T = trace Y Y T = trace { G} where G = Y T Y is the n n Gram matrix. Thus: J is quadratic in Y, and positive semidefnite, hence convex. Also: J is linear in G. Question: Are there other convex functions in Y that behave similarly?

7 Optimization Criteria Equivalent forms: J(Y ) = trace { } { } Y (D W )Y T = trace Y Y T = trace { G} where G = Y T Y is the n n Gram matrix. Thus: J is quadratic in Y, and positive semidefnite, hence convex. Also: J is linear in G. Question: Are there other convex functions in Y that behave similarly? Answer: Yes! Examples: J(y 1,, y n ) = W i,j y i y j 1 i,j n

8 Optimization Criteria Equivalent forms: J(Y ) = trace { } { } Y (D W )Y T = trace Y Y T = trace { G} where G = Y T Y is the n n Gram matrix. Thus: J is quadratic in Y, and positive semidefnite, hence convex. Also: J is linear in G. Question: Are there other convex functions in Y that behave similarly? Answer: Yes! Examples: J(y 1,, y n ) = W i,j y i y j J(y 1,, y n ) = 1 i,j n 1 i,j n W i,j y i y j p 1/p, p 1

9 Constraints Absent any constraint, minimize trace {Y } Y T has solution Y = 0. To avoid this trivial solution, we impose a normalization constraint. Choices: YY T = I d YDY T = I d What does this mean? n y k yk T = I d Parseval frame k=1 n d k y k yk T = I d Parseval weighted frame k=1

10 The Optimization Problem Combining one criterion with one constraint: minimize trace {Y } Y (LE) : T subject to YDY T = I d called the Laplacian Eigenmap problem. Alternative problem: (UnLE) : minimize subject to trace {Y } Y T YY T = I d called the unnormalized Laplacian eigenmap problem.

11 The optimization problem How to solve the Laplacian eigenmap problem: minimize trace {Y } Y (LE) : T subject to YDY T = I d First note the problem is not convex, because of the equality constraint. How to make it convex? How to solve? 1. First absorb the scaling D into the solution: Ỹ = YD 1/2 Problem becomes: minimize trace {Ỹ } D 1/2 D 1/2 Ỹ T = trace {Ỹ Ỹ } T subject to Ỹ Ỹ T = I d

12 The optimization problem 2. Consider the optimization problem for P: { } minimize trace P subject to P = P T 0 P I n trace(p) = d Proposition Claims: A. The above optimization problem is a convex SDP. B. At optimum: P = Ỹ T Ỹ.

13 Eigenproblem The optimum solutions of the (LE) and (UnLE) problems are given by appropriate eigenvectors: minimize trace {Ỹ Ỹ } T subject to Ỹ Ỹ T = I d Solution: Ỹ = e T 1. ed T, e k = λ k e k where 0 = λ 1 λ d are the smallest d eigenvalues, and e k = 1 are the normalized eigenvectors.

14 Generalized Eigenproblem (LE) : minimize trace {Y } Y T subject to YDY T Y = Ỹ D 1/2 = I d the rows of Ỹ are eigenvectors of the normalized Laplacian e k = λ k e k. Let f k be the (transpose) rows of Y : Y = f T 1. f T d, f k = D 1/2 e k Thus: D 1/2 f k = λ k D 1/2 f k, or: D 1/2 D 1/2 f k = λ k Df k, or: f k = λ k Df k This is called generalized eigenproblem associated to (, D).

15 Eigenproblem Consider the unnormalized Laplacian eigenmap problem: minimize trace {Y } Y (UnLE) : T subject to YY T. = I d The solution Y unle is the d n matrix whose rows are eigenvectors of the unnormalized Laplacian = D W, g k = µ k g k, g k = 1, 0 = µ 1 µ d, and Y unle = g T 1. gd T.

16 What eigenspace to choose? In most implementations one skips the eigenvectors associated to 0 eigenvalue. Why? In the unnormalized case, g 1 = 1 n [1, 1,, 1] T, hence no new information. In your class projects, skip the bottom eigenvector. This fact is explicitely stated in the problem.

17 Embedding Algorithm Algorithm () Input: Weight matrix W, target dimension d 1 Construct the diagonal matrix D = diag(d ii ) 1 i n, where D ii = n k=1 W i,k. 2 Construct the normalized Laplacian = I D 1/2 WD 1/2. 3 Compute the bottom d + 1 eigenvectors e 1,, e d+1, e k = λ k e k, 0 = λ 1 λ d+1.

18 Embedding Algorithm-cont s Algorithm ( - cont d) 4 Construct the d n matrix Y, e 2 Y =. D 1/2 e d+1 5 The new geometric graph is obtained by converting the columns of Y into n d-dimensional vectors: [ y 1 y n ] = Y Output: Set of points {y 1, y 2,, y n } R d.

19 Example see: rvbalan/teaching/amsc663fall2010/ PROJECTS/P5/index.html

20 References B. Bollobás, Graph Theory. An Introductory Course, Springer-Verlag (25), (2002). S. Boyd, L. Vandenberghe, Convex Optimization, available online at: boyd/cvxbook/ F. Chung, Spectral Graph Theory, AMS F. Chung, L. Lu, The average distances in random graphs with given expected degrees, Proc. Nat.Acad.Sci F. Chung, L. Lu, V. Vu, The spectra of random graphs with Given Expected Degrees, Internet Math. 1(3), (2004). R. Diestel, Graph Theory, 3rd Edition, Springer-Verlag P. Erdös, A. Rényi, On The Evolution of Random Graphs

21 G. Grimmett, Probability on Graphs. Random Processes on Graphs and Lattices, Cambridge Press C. Hoffman, M. Kahle, E. Paquette, Spectral Gap of Random Graphs and Applications to Random Topology, arxiv: [math.co] 17 Sept A. Javanmard, A. Montanari, Localization from Incomplete Noisy Distance Measurements, arxiv: , Nov. 2012; also ISIT J. Leskovec, J. Kleinberg, C. Faloutsos, Graph Evolution: Densification and Shrinking Diameters, ACM Trans. on Knowl.Disc.Data, 1(1) 2007.