The non-backtracking operator

Transcription

1 The non-backtracking operator Florent Krzakala LPS, Ecole Normale Supérieure in collaboration with Paris: L. Zdeborova, A. Saade Rome: A. Decelle Würzburg: J. Reichardt Santa Fe: C. Moore, P. Zhang Berkeley: E. Mossel, A. Sly, J. Neeman

2 Community detection in network Stochastic block modeling (statistics) Data Clustering (machine learning) Graph-based machine learning (inference) Planted constraint satisfaction models Given the graph, find the labeling! Online communities Word adjacency networks Food webs Metabolic networks Protein-protein interaction networks Financial market (sectors)...

3 This is a hard problem Adjacency matrix What you have: The data What you want: The structure

4 This is a hard problem Karate club 77 The internet Protein network and it is getting harder!

5 Algorithms for clustering Hundreds of different methods - for a review e.g. S. Fortunato, Physics Reports, Spectral clustering - the state of art clustering method in machine learning. Associate a matrix to the network (adjacency, random walk, Laplacian, etc.), compute its top eigenvalues. The corresponding eigenvectors encode the clusters in a visible way.

6 This talk Inference using the stochastic block model Spectral algorithm(s)

7 Stochastic block model Generate a random network as follows: q groups, N nodes n a proportion of nodes in group p ab = c ab a = 1,..., q probability that an edge present N between node from group a and another from group b n 1 = 7/12 n 2 = 5/12 p 11 = p 22 = 0.39 p 12 = p 21 = 0.14

8 Optimal inference in SBM (when parameters of the model known) P ({q i })= NY n qi P (A ij {q i })= Y i6=j p A ij q i q j (1 p qi q j ) 1 A ij i=1 P ({q i } A ij )= 1 Z P ({q i})p (A ij {q i }) n a p ab proportion of nodes in group a proportion of edge between groups a and b q i 2{1,...,q} an assignment of colors Includes ALL available information about the signal

9 Optimal algorithm in SBM (for a statistician) Posterior probability distribution P ({s i } A ij )= 1 Z P ({s i})p (A ij {s i }) Marginal probabilities µ(s i ) X {s j } j6=i P ({s j } j6=i,s i A ij ) Maximize number of correctly assigned nodes s i = argmax si µ(s i )

10 Optimal algorithm in SBM (for a statistical physicist) Potts spin variables s i 2 {1,...,q} Boltzmann measure P ({s i })= 1 Z e H({s i}) Hamiltonian: Potts glass in field on the Nishimori line H({s i })= NX log n si i=1 i6=j magnetic field X Aij log p si s j +(1 A ij ) log (1 p si s j ) pair-wise interactions Local magnetization s i = argmax si µ(s i )

11 How to compute marginals and Z? A #P-hard problem in general MCMC: (Monte Carlo Markov chain, Gibbs sampling): generic, but potentially a (very) large equilibration time Variational mean field: convergence problem and (sometimes) a crude approximation. Belief propagation/cavity method: fast, hard to control, generally better than mean field, exact for large networks generated by the model.

12 Results (in a nutshell) Q = max P N i=1 (t i ),s i 1 1/q /N 1/q Overlap of the optimal estimation with the true labeling Q N=100k, BP N=70k, MC 0.6 q=4, c= ferromagnet detectable paramagnet undetectable ε= c out /c in

13 Results (in a nutshell) overlap c d init. planted init. random q=5,c in =0, N=100k c s paramagnet 0 spinodal region ferromagnet dis-assortative network 5 colors c

14 Algorithmic consequences c d c s c<c c impossible init. ordered init. random log Z pl -log Z ran q=5,c in =0, N=100k log Z pl -log Z ran 0 c c c c in c out = q p c s c c <c<c s c s <c possible but hard easy 0

15 Rigorous results for q=2 For q=2 we find that the information theoretic thresholds equals the algorithmic one c in c out = q p c s Proven! Communutities are undetectable bellow cs and easy to find beyond cs The undetectable regime (Mossel, Neeman, Sly 12) : Planted graphs are essentially Erdos-Renyi random graphs in the undetectable regime. The detectable regime (Massoulie 13)

16 This talk Inference using the stochastic block model Spectral algorithm(s)

17 Spectral clustering Compute k largest eigenvalues and their eigenvectors for a matrix associated to the graph. Cluster these eigenvectors, e.g. using k-means. For 2 groups - signs of the 2nd eigenvector. Matrices: Adjacency Laplacian Random walk Modularity A ij L ij = d i i,j Q ij = A ij d i M ij = A ij A ij d i d j 2M

18 Adjacency matrix for dense graphs Largest eigenvalue is the average degree, and the eigenvector is trivial. It exists an eigenvector correlated with the communities, with eigenvalue c = cin cout 2 cin + cout + cin cout 1 p 4c The bulk follows Wigner s semi circle law: P ( ) = 2 c If cin 2 p cout > 2 c the informative eigenvalue is out of the bulk Adjacency is optimal for dense graphs! If it fails, any thing else will Nadakuditi & Newman 12

19 Performance of spectral clustering Overlap B Modularity Random Walk A L BP q =2 n a =1/2 c =3 N = 10 5 c in = c aa c out = c a6=b c in c out = q p c c in c out Q = max P N i=1 (t i ),q i 1 1/q /N 1/q

20 Why the suboptimality? Wigner not good for sparse G P ( )= 1 p 4c 2 2 c The edge of the spectrum of sparse graphs is spoiled by nodes of large degree. ER graphs, largest degree ~ log(n)/log(log(n)). How to correct this? Remove largest degrees? Not good enough - loosing information.

21 Spectral Redemption Can the optimal performance (phase transition) be matched by a spectral method?

22 The non-backtracking matrix M edges, define matrix B on directed edges, i.e. 2M x 2M matrix as follows B i!j,k!l =1 if j = k, i 6= l B i!j,k!l = 0 otherwise The Hashimoto ( 89) edge adjacency operator used to evaluate the Ihara zeta function Directed walk similar to a message passing algorithm (i.e. Belief propagation)

23 The non-backtracking matrix µ Not a random sparse real non-symmetric matrix, elements correlated! adjacency matrix spectrum

24 Performance of spectral clustering B Modularity Random Walk A L BP Overlap c in c out

25 Properties of the non-backtracking matrix Eigenvalues of B, and sums of eigenvector elements are given by a 2N x 2N matrix. B 0 = Characteristic polynomial (Bass 92 formula) Largest eigenvalue c. c in c out = q p c 0 D 1 1 A v j in = X i2@j det µ 2 1 µa +(D 1) =0 v i!j If the 2nd eigenvalue is informative and > p c All others are inside circle p c (we believed, but proved only in density) Number of eigenvalues outside the circle = number of communities! Complete proof now available Massoulie, Lelarge and Bordenave

26 Spectra of some real networks

27 Spectrum of the non-backtracking matrix

28 Spectrum of the non-backtracking matrix (with some physics hocus-pocus ) det D za (1 z 2 )1 = (z) = 1 2N 2NX i=1 2NY i (z i) (z µ) = z(z µ) 1 (z i). 1 (z) =lim!0 2 z@ z log det M M (z,a) = 1 i(d za (1 z 2 )1) i(d za (1 z 2 )1) 1 (z) = 1 lim!0 2 z@ z log Z

29 Spectrum of the non-backtracking matrix (with some physics hocus-pocus )

30 Spectrum of the non-backtracking matrix (with some physics hocus-pocus ) Non-localized localized

31 The non-backtracking operator Allows spectral clustering to match information theoretic performances Trivial to implement: spectrum of 2N x 2N matrix. B 0 = 0 D 1 1 A Parameter free: No need for learning Avoid the usual BP convergence problems on non-tree like graph.

32 The non-backtracking operator Perspectives: Can we compute Full spectrum rigorously? I believe we have only scratched the surface of the potential of this approach Spectral method for matrix factorization? Censored block model (Abbe et al)? Percolation?

33 Percolation Dense network: The critical occupation probability for percolation on a dense network is equal to the reciprocal of the leading eigenvalue of the adjacency matrix (Bollobas et al 10) ANY network: The critical occupation probability for percolation on a dense network is bounded to the reciprocal of the leading eigenvalue of the non-backtracking matrix (Zdeborova et al 14) (and exact for sparse tree-like graph and dense graphes as well..)

34 This talk is based on: A. Decelle, FK, C. Moore, L. Zdeborová, Phase transition in the detection of modules in sparse networks, Phys. Rev. Lett. 107, (2011). A. Decelle, FK, C. Moore, L. Zdeborová, Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications, Phys. Rev. E 84, (2011). P. Zhang, FK, J. Reichardt, L. Zdeborová, Comparative Study for Inference of Hidden Classes in Stochastic Block Models, J. Stat. Mech. (2012) P12021 FK, C. Moore, E. Mossel, J. Neeman, A. Sly, L. Zdeborová, P. Zhang, Spectral clustering of Sparse Networks, Proc. Nat. Acad. Sci. (2013) A. Saade, FK & L. Zdeborová, The spectrum of the non-backtracking matrix EPL (2014) A. Saade, FK & L. Zdeborová, Clustering with the Bethe Hessian NIPS 2014 Implementations available at: Thank you for your attention!