LecXX5

LecXX5 - constant values in 2 nd eigenvector COT 6936 4...

Info iconThis preview shows pages 1–3. Sign up to view the full content.

View Full Document Right Arrow Icon
1 4/6/10 COT 6936 1 COT 6936: Topics in Algorithms Giri Narasimhan ECS 254A / EC 2443; Phone: x3748 [email protected] http://www.cs.fiu.edu/~giri/teach/COT6936_S10.html https://online.cis.fiu.edu/portal/course/view.php?id=427 Spectral Methods ± Graph Connectivity problems ± Google Page Rank ± Graph Partitioning problems ± Clustering (even linearly non-separable case) ± Markov Chain Mixing problems ± Random walks in graphs 4/6/10 COT 6936 2 Matrices and Eigenvalues ± Array of values ± Linear Transformation ± Eigenvalues and Eigenvectors ± Ax = ± x ± Under transformation A, eigenvectors only experience change in magnitude, not direction ± A = Q ² Q -1 4/6/10 COT 6936 3
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full Document Right Arrow Icon
2 Graph Bisection ± Construct adjacency matrix A ± Construct Laplacian L = D – A ± L is positive semi-definite (PSD); has non-neg eigenvalues; has smallest eigenvalue = 0 ± Second eigenvector provides information about bisection. ± Signs of 2 nd eigenvector give a good bisection ± Extreme case : Connected components have
Background image of page 2
Background image of page 3
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: constant values in 2 nd eigenvector 4/6/10 COT 6936 4 Graph Bisection (Continued) ± Eigenvalues indicate strength of bisection ± How to get bisections with n/2 vertices? ± Use median value in second eigenvector ± How to get k partitions? ± Perform bisections recursively ± Use more eigenvectors 4/6/10 COT 6936 5 Spectral Clustering: Strategy ± Given data points and a distance function, construct a weighted graph ± Let A be its adjacency matrix; let D be diagonal matrix with degrees along diagonal ± Construct Laplacian L ( PSD , non-neg eigenv .) ± Unnormalized: L = D – A ± Normalized symmetric: L = D-1/2 LD 1/2 ± Random Walk: L = D-1 L ± Matrix L k has cols = first k eigenvectors of L ± Cluster rows of L k . 4/6/10 COT 6936 6 3 Spectral Clustering ± Need distance measure (need not be a metric), i.e., triangle inequality not needed ± Not Model-based ± Global method ± Turns discrete problem into continuous 4/6/10 COT 6936 7...
View Full Document

{[ snackBarMessage ]}

Page1 / 3

LecXX5 - constant values in 2 nd eigenvector COT 6936 4...

This preview shows document pages 1 - 3. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online