Normalized Cuts Are Approximately Inverse Exit Times

Normalized cut is a widely used measure of separation between clusters in a graph. In this paper we provide a novel probabilistic perspective on this measure. We show that for a partition of a graph into two weakly connected sets, $V=A\uplus B$, the multiway normalized cut is approximately $MNcut \approx 1/\tau_{A\to B}+1/\tau_{B\to A}$, where $\tau_{A\to B}$ is the unidirectional characteristic exit time of a random walk from subset $A$ to subset $B$. Using matrix perturbation theory, we show that this approximation is exact to first order in the connection strength between the two subsets $A$ and $B$, and we derive an explicit bound for the approximation error. Our result implies that for a Markov chain composed of a reversible subset $A$ that is weakly connected to an absorbing subset $B$, the easy-to-compute normalized cut measure is a reliable proxy for the chain's spectral gap.

[1]  Santosh S. Vempala,et al.  On clusterings: Good, bad and spectral , 2004, JACM.

[2]  Stéphane Lafon,et al.  Diffusion maps , 2006 .

[3]  C. Ding,et al.  On the Equivalence of Nonnegative Matrix Factorization and K-means - Spectral Clustering , 2005 .

[4]  Jitendra Malik,et al.  Normalized cuts and image segmentation , 1997, Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition.

[5]  Carl D. Meyer,et al.  Stochastic Complementation, Uncoupling Markov Chains, and the Theory of Nearly Reducible Systems , 1989, SIAM Rev..

[6]  Michael I. Jordan,et al.  On Spectral Clustering: Analysis and an algorithm , 2001, NIPS.

[7]  Liang Xu,et al.  Regularized spectral learning , 2005, AISTATS.

[8]  Ronald R. Coifman,et al.  Diffusion Maps, Spectral Clustering and Eigenfunctions of Fokker-Planck Operators , 2005, NIPS.

[9]  Jianbo Shi,et al.  Multiclass spectral clustering , 2003, Proceedings Ninth IEEE International Conference on Computer Vision.

[10]  B. Matkowsky,et al.  Eigenvalues of the Fokker–Planck Operator and the Approach to Equilibrium for Diffusions in Potential Fields , 1981 .

[11]  Jianbo Shi,et al.  A Random Walks View of Spectral Segmentation , 2001, AISTATS.

[12]  G. Stewart,et al.  Matrix Perturbation Theory , 1990 .

[13]  Michael I. Jordan,et al.  Learning Spectral Clustering, With Application To Speech Separation , 2006, J. Mach. Learn. Res..

[14]  P. Deuflharda,et al.  Identification of almost invariant aggregates in reversible nearly uncoupled Markov chains , 2000 .

[15]  Inderjit S. Dhillon,et al.  Kernel k-means: spectral clustering and normalized cuts , 2004, KDD.

[16]  Moshe Haviv,et al.  On Series Expansions and Stochastic Matrices , 1993, SIAM J. Matrix Anal. Appl..

[17]  K. Avrachenkov,et al.  Singular perturbations of Markov chains and decision processes , 2002 .

[18]  Meirav Galun,et al.  Fundamental Limitations of Spectral Clustering , 2006, NIPS.