Spectral Graph Theory Tools. Analysis of Complex Networks

Transcription

1 Spectral Graph Theory Tools for the Department of Mathematics and Computer Science Emory University Atlanta, GA 30322, USA

2 Acknowledgments Christine Klymko (Emory) Ernesto Estrada (Strathclyde, UK) Support: NSF DMS Papers and preprints available on web page

3 Outline Complex networks: motivation and background Graph spectra: a brief introduction Spectral centrality and communicability measures Clustering coefficients

4 Complex networks: motivation and background Complex networks provide models for physical, biological, engineered or social systems (e.g., molecular structure, gene and protein interaction, food webs, transportation networks, power grids, social networks,...). Graph analysis provides quantitative tools for the study of complex networks. Techniques from spectral graph theory, linear and multilinear algebra, probability, approximation theory, etc. play a major role. Network science today is a vast multidisciplinary field. Important early work was done by social scientists: sociologists, anthropologists, experimental psychologists, economists and even bibliometrists. More recently, physicists, computer scientists and applied mathematicians have made major contributions.

5 Basic references Some classic early references: L. Katz, A new status index derived from sociometric analysis, Psychometrika, 18 (1953), pp A. Rapoport, Mathematical models of social interaction, in Handbook of Mathematical Psychology, vol. 2, pp Wiley, New York, D. J. de Solla Price, Networks of scientific papers, Science, 149 (1965), pp S. Milgram, The small world problem, Psychology Today, 2 (1967), pp J. Travers and S. Milgram, An experimental study of the small world problem, Sociometry, 32 (1969), pp

6 Basic references (cont.) The field exploded in the late 1990s due to several breakthroughs by physicists, applied mathematicians and computer scientists. This helped make the field more quantitative and mathematically sophisticated. Landmark papers include D. J. Watts and S. H. Strogatz, Collective dynamics of small-world networks, Nature, 393 (1998), pp L.-A. Barabási and R. Albert, Emergence of scaling in random networks, Science, 386 (1999), pp M. E. J. Newman, Models of the small world, J. Stat. Phys., 101 (2000), pp J. Kleinberg, Navigation in a small world, Nature, 406 (2000), p R. Albert and L.-A. Barabási, Statistical mechanics of complex networks, Rev. Modern Phys., 74 (2002), pp

7 Basic references (cont.) Recent systematic accounts include U. Brandes and T. Erlebach (Eds.), Network Analysis. Methodological Foundations, Springer, LNCS 3418, F. Chung and L. Lu, Complex Graphs and Networks, American Mathematical Society, A. Barrat, M. Barthelemy, and A. Vespignani, Dynamical Processes on Complex Networks, Cambridge University Press, M. E. J. Newman, Networks. An Introduction, Oxford University Press, P. Van Mieghem, Graph Spectra for Complex Networks, Cambridge University Press, E. Estrada, The Structure of Complex Networks, Oxford University Press, 2011.

8 Basic references (cont.) Two new interdisciplinary journals in 2013: Journal of Complex Networks (Oxford University Press) Network Science (Cambridge University Press)

9 Complex networks: what are they? But what exactly is a complex network? Unfortunately, no precise definition exists. It is easier to say which graphs are not complex networks. Regular lattices are not considered complex networks, and neither are completely random graphs such as the Gilbert or Erdös Rényi models. Random graphs are, however, considered by many to be useful as toy models for complex networks.

10 Regular lattice: not a complex network!

11 Star graph: also not a complex network!

12 Erdös Rényi graph: also not a complex network!

13 Some features of complex networks Some of the attributes typical of many real-world complex networks are: Scale-free (power law degree distribution) Small-world (small graph diameter) Highly clustered (many triangles, hubs...) Hierarchical Rich in motifs Self-similar Caveat: there are important examples of real-world complex networks lacking one or more of these attributes.

14 Example of complex network: B A model

15 Example of complex network: Golub collaboration graph

16 Example of complex network: Erdös collaboration graph

17 Example of complex network: PPI network of Saccharomyces cerevisiae (beer yeast)

18 Example of complex network: social network of injecting drug users in Colorado Springs, CO

19 Example of complex network: the Internet

20 Example of (directed) complex network: a food web

21 Network analysis Basic questions about network structure include centrality, communicability and community detection issues: Which are the most important nodes? Network connectivity (vulnerability) Lethality in PPI networks Author centrality in collaboration networks How do disturbances spread in a network? Spreading of epidemics, rumors, fads,... Routing of messages; returnability How to identify community structures in a network? Clustering, transitivity Partitioning

22 Basic definitions Notation: G =(V, E) is a connected simple graph with N = V nodes and m = E edges (i.e., G is unweighted, undirected and has no loops) A =[a ij ] R N N is the associated adjacency matrix: a ij = 1 if nodes i and j are adjacent, a ij = 0 otherwise a ii = 0, 1 i N A is symmetric For a node i, define its degree d i := N j=1 a ij (number of immediate neighbors in G) Graph Laplacian: L := D A where D = diag (d 1,...,d N ) Some modifications necessary for digraphs (directed graphs)

23 Network spectra For any matrix A R N N, the (generally complex) number λ is an eigenvalue of A if there exists a vector x = 0 such that Ax = λx. We say that x is an eigenvector of A, associated with the eigenvalue λ. Note that eigenvectors are defined up to nonzero scalar multiples. Eigenvectors are often normalized to have unit length: x 2 = x x x N 2 =1.

24 Network spectra (cont.) The eigenvalues are the roots of the characteristic equation det(a λi N )=0, a polynomial equation of degree N. This can be seen from the fact that the homogeneous system of linear equations Ax λx =(A λi N )x = 0 can have non-zero solutions if and only if the matrix A λi N is singular.

25 Network spectra (cont.) The Fundamental Theorem of Algebra ensures that any N N matrix A has N eigenvalues, counted with their multiplicities. The set consisting of the eigenvalues of A: is called the spectrum of A. σ(a) :={λ 1,λ 2,...,λ N } Recall that a matrix A =[a ij ] is symmetric if A = A T ; that is, a ij = a ji for all i, j =1:N. The eigenvalues of a real symmetric matrix are all real. Eigenvectors corresponding to distinct eigenvalues of a symmetric matrix are necessarily orthogonal.

26 Network spectra (cont.) A general (nonsymmetric) matrix may have complex eigenvalues. If A is real, the complex eigenvalues appear in conjugate pairs. Hence, a real matrix of odd order always has at least one real eigenvalue. An example of a real matrix with no real eigenvalues is given by 0 1 A = 1 0 The adjacency matrix of an undirected network is always symmetric, hence its eigenvalues are all real. The adjacency matrix of a directed network, on the other hand, is typically nonsymmetric and will have complex (non-real) eigenvalues in general. The spectrum of a network is the spectrum of the corresponding adjacency matrix.

27 Network spectra (cont.) Perron Frobenius Theorem: Let A R N N be a matrix with nonnegative entries. If A is irreducible, the spectral radius of A, defined as ρ(a) =max{ λ : λ σ(a)} is a simple eigenvalue of A. Moreover, there exists a unique positive eigenvector x associated with this eigenvalue: Ax = ρ(a)x, x =(x i ), x i > 0 i =1:N. This theorem is of fundamental importance, and it provides the basis for the notion of eigenvector centrality and for Google s celebrated PageRank algorithm. It is also probably behind Facebook s new Graph Search tool (cf. Mark Zuckerberg s announcement yesterday).

28 Network spectra (cont.) For an undirected network, the spectrum of the graph Laplacian L = D A also plays an important role. To avoid confusion, we call σ(l) the Laplacian spectrum of the network. The eigenvalues of L are all real and nonnegative, and since L 1 = 0, where 1 denotes the vector of all ones, L is singular. Hence, 0 σ(l). Theorem. The multiplicity of 0 as an eigenvalue of L coincides with the number of connected components of the network. Corollary. For a connected network, the null space of the graph Laplacian is 1-dimensional. Thus, rank(l) = N 1.

29 Network spectra (cont.) The eigenvector of L associated with the smallest nonzero eigenvalue is called the Fiedler vector of the network. Since this eigenvector must be orthogonal to the vector of all ones, it must contain both positive and negative entries. There exist elegant graph partitioning algorithms that assign nodes to different subgraphs based on the sign of the entries of the Fiedler vector. These methods, however, tend to work well only in the case of fairly regular graphs. In general, partitioning complex graphs (scale-free graphs in particular) is very difficult.

30 Example: PPI network of Saccharomyces cerevisiae nz = Adjacency matrix, N = 2224, m = 6609.

31 Example: PPI network of Saccharomyces cerevisiae nz = Same, reordered with Reverse Cuthill McKee

32 Example: PPI network of Saccharomyces cerevisiae Degree distribution (d min = 1, d max = 64)

33 Example: PPI network of Saccharomyces cerevisiae Eigenvalues

34 Example: Intravenous drug users network nz = 4024 Adjacency matrix, N = 616, m = 2012.

35 Example: Intravenous drug users network nz = 4024 Same, reordered with Reverse Cuthill McKee

36 Example: Intravenous drug users network Degree distribution (d min = 1, d max = 58)

37 Example: Intravenous drug users network Eigenvalues

38 Centrality measures There are dozens of different definitions of centrality for nodes in a graph. The simplest is degree centrality, which is just the degree d i of node i. This does not take into account the importance of the nodes a given nodes is connected to only their number. A popular notion of centrality is betweenness centrality, defined for any node i V as c B (i) := δ jk (i), j=i k=i where δ jk (i) is the fraction of all shortest paths in the graph between nodes j and k which contain node i: δ jk (i) := # of shortest paths between j, k containing i # of shortest paths between j, k.

39 Centrality measures (cont.) Betweenness centrality assumes that all communication in the network takes place via shortest paths, but this is often not the case. This observation has motivated a number of alternative definitions of centrality, which aim at taking into account the global structure of the network and the fact that all walks between pairs of nodes should be considered, not just shortest paths. We mention here that for directed graphs it is sometimes necessary to consider both hubs and authorities (Ex.: Jon Kleinberg s HITS algorithm). I hope Christine Klymko will give a presentation on this.

40 Spectral centrality measures Bonacich s eigenvector centrality uses the entries of the dominant eigenvector (that is, the eigenvector corresponding to the spectral radius of the adjacency matrix) to rank the nodes in the network in order of importance: the larger x i is, the more important node i is considered to be. By the Perron Frobenius Theorem, this eigenvector is unique provided the network is connected. The underlying idea is that a node is important if it is linked to many important nodes. This circular definition corresponds to the fixed-point iteration x k+1 = Ax k, k =0, 1,... which converges, under mild conditions, to the dominant eigenvector of A as k. The rate of convergence depends on the spectral gap γ = λ 1 λ 2. The larger γ, the faster the convergence.

41 Spectral centrality measures (cont.) Google s PageRank algorithm is a variant of eigenvector centrality, applied to the (directed) graph representing web pages (documents), with hyperlinks between pages playing the role of directed edges. Since the WWW graph is not connected, some tweaks are needed to have a unique PageRank eigenvector. The entries of this vector have a probabilistic interpretation in terms of random walks on the web graph, and their sizes determine the ranking of web pages. The PageRank vector is the unique stationary probability distribution of the Markov chain corresponding to the random walk on the Web graph.

42 Spectral centrality measures (cont.) When a user enters a term or phrase into Google s search engine, links to pages containing those terms are returned in order of importance. PageRank is (was?) one of the criteria Google uses to rank pages. Since hardly anybody looks past the few (5-10) top results, it s very important to get these top pages right. Google s unsurpassed ability to do this gave it an early edge on the competition in the late 1990s. Computation of PageRank is quite expensive due to the enormous size of the WWW graph.

43 Subgraph centrality Subgraph centrality (Estrada & Rodríguez-Velásquez, 2005) measures the centrality of a node by taking into account the number of subgraphs the node participates in. This is done by counting, for all k =1, 2,... the number of closed walks in G starting and ending at node i, with longer walks being penalized (given a smaller weight).

44 Subgraph centrality Recall that (A k ) ii = # of closed walks of length k based at node i, (A k ) ij = # of walks of length k that connect nodes i and j. Using 1/k! as weights leads to the notion of subgraph centrality: SC(i) = I + A + 1 2! A ! A3 + ii = [e A ] ii Subgraph centrality has been used successfully in various settings, especially proteomics and neuroscience. Note: the additional term of order k = 0 does not alter the ranking of nodes in terms of subgraph centrality.

45 Katz centrality Of course different weights can be used, leading to different matrix functions, such as the resolvent (Katz, 1953): (I ca) 1 = I + ca + c 2 A 2 +, 0 < c < 1/λ 1. Also, in the case of a directed network one can use the solution vectors of the linear systems (I ca)x = 1 and (I ca T )y = 1 to rank hubs and authorities. These are the row and column sums of the matrix resolvent (I ca) 1, respectively. Similarly, one can use the row and column sums of e A to rank nodes. This can be done without computing the entries of e A, and is much faster than computing the diagonal entries of e A.

46 Comparing centrality measures Suppose A is the adjacency matrix of an undirected, connected network. Let λ 1 >λ 2 λ 3... λ N be the eigenvalues of A, and let x k be a normalized eigenvector corresponding to λ k. Then and therefore e A = SC(i) = N e λ k x k x T k k=1 N e λ k xki 2, k=1 where x ki denotes the ith entry of x k. Similar spectral formulas exist for other centrality measures based on matrix functions (e.g., Katz).

47 Comparing centrality measures (cont.) It is clear that when the spectral gap is sufficiently large (λ 1 λ 2 ), the centrality ranking is essentially determined by the dominant eigenvector x 1, since the contributions to the centrality scores from the remaining eigenvalues/vectors becomes negligible, in relative terms. Hence, in this case all these spectral centrality measures tend to agree, especially on the top-ranked nodes, and they all reduce to eigenvector centrality. Degree centrality is also known to be strongly correlated with eigenvector centrality in this case. In contrast, the various measures can give significantly different results when the spectral gap is small. In this case, going beyond the dominant eigenvector can result in significant improvements.

48 More matrix functions Other matrix functions of interest are cosh(a) and sinh(a), which correspond to considering only walks of even and odd length, respectively. Because in a bipartite graph Tr (sinh(a)) = 0, the quantity B(G) := Tr (cosh(a)) Tr (e A ) provides a measure of how close a graph is to being bipartite. Here Tr( ) is the trace (sum of diagonal entries) of a matrix. Hyperbolic matrix functions are also used to define the notion of returnability in digraphs (Estrada & Hatano, 2009).

49 Communicability Communicability measures how easy it is to send a message from node i to node j in a graph. It is defined as (Estrada & Hatano, 2008): C(i, j) = [e A ] ij = I + A + 1 2! A ! A3 + ij weighted sum of walks joining nodes i and j. As before, other power series expansions (weights) can be used, leading to different functions of A. For a large graph, computing all communicabilities C(i, j) (i = j) is prohibitively expensive. Instead, averages are often used as in statistical physics.

50 Communicability The average communicability of a node is defined as C(i) := 1 C(i, j) = 1 N 1 N 1 j=i [e A ] ij. Communicability functions can be used to study the spreading of diseases (or rumors) and to identify bottlenecks in networks. See Estrada, Hatano and B. (Phys. Reports, 2012) for a review. Recently, community detection algorithms based on various communicability measures have been developed (Fortunato, 2010; Estrada, 2011). j=i

51 Clustering coefficients A clustering coefficient measures the degree to which the nodes in a network tend to cluster together. For a node i with degree d i, it is defined as 2 i CC(i) = d i (d i 1) where i is the number of triangles in G having node i as one of its vertices. The clustering coefficient of a graph G is defined as the sum of the clustering coefficients over all the nodes of degree 2. Many real world small-world networks, and particularly social networks, tend to have high clustering coefficient (relative to random networks).

52 Clustering coefficients (cont.) The number of triangles in G that a node participates in is given by i = 1 2 [A3 ] ii, while the total number of triangles in G is given by (G) = 1 6 Tr (A3 ). Computing clustering coefficients for a graph G reduces to computing diagonal entries of A 3. We note for many networks, A 3 is a rather dense matrix. For example, for the PPI network of beer yeast the percentage of nonzero entries in A 3 is about 19%, compared to 0.27% for A.