The PageRank of a graph Fan Chung University of California, San Diego What is PageRank? •Ranking the vertices of a graph •Partially ordered sets + graph theory A new graph invariant for dealing with: -Applications --web search,ITdata sets, xxxxxxxxxxxxxpartitioning algorithms,… -Theory --- correlations among vertices fan Outline of the talk • Motivation • Define PageRank • Local cuts and the Cheeger constant • Four versions of the Cheeger inequality using: eigenvectors random walks PageRank heat kernel • Four partitioning algorithms • Green’s functions and hitting time What is PageRank? What is Rank? What is PageRank? PageRank is defined on any graph. An induced subgraph of the collaboration graph with authors of Erdös number ≤ 2. A subgraph of the Hollywood graph. A subgraph of a BGP graph The Octopus graph Yahoo IM graph Reid Andersen 2005 Graph Theory has 250 years of history. Leonhard Euler 1707-1783 The Bridges of Königsburg Is it possible to walk over every bridge once and only once? Geometric graphs Topological graphs Algebraic graphs General graphs Massive data Massive graphs • WWW-graphs • Call graphs • Acquaintance graphs • Graphs from any data a.base protein interaction network Jawoong Jeong Big and bigger graphs New directions in graph theory Many basic questions: • Correlation among vertices? • The `geometry’ of a network ? distance, flow, cut, … • Quantitative analysis? eigenvalues, rapid mixing, … • Local versus global? Google’s answer: The definition for PageRank? A measure for the “importance” of a website x1 R( x14 x79 x785 ) x2 R( x1002 x3225 x9883 x30027 ) The “importance” of a website is proportional to the sum of the importance of all the sites that link to it. A solution for the “importance” of a website x1 ( x14 x79 x785 ) x2 ( x1002 x3225 x9883 x30027 ) Solve n xi aij x j j 1 for x ( x1 , x2 , , xn ) A solution for the “importance” of a website x1 ( x14 x79 x785 ) x2 ( x1002 x3225 x9883 x30027 ) Solve n xi aij x j for x ( x1 , x2 , , xn ) j 1 x=ρAx A aij nn Adjacency matrix A solution for the “importance” of a website x1 ( x14 x79 x785 ) x2 ( x1002 x3225 x9883 x30027 ) Solve n xi aij x j j 1 x=ρAx for x ( x1 , x2 , , xn ) Graph models (undirected) graphs directed graphs weighted graphs Graph models (undirected) graphs directed graphs weighted graphs Graph models (undirected) graphs directed graphs weighted graphs 2.3 1.2 1.5 1.1 1 2 3.3 2.8 \ 1.5 In a directed graph, there are two types of “importance”: authority hub Jon Kleinberg 1998 Two types of the “importance” of a website Importance as Authorities : Importance as Hubs : Solve and x=rAy T x = rs A A x T y = rs A A y x ( x1 , x2 , , xn ) y ( y1 , y2 , , ym ) T y=sAx Eigenvalue problem for n x n matrix:. n ≈ 30 billion websites Hard to compute eigenvalues Even harder to compute eigenvectors In the old days, compute for a given (whole) graph. In reality, can only afford to compute “locally”. A traditional algorithm Input: a given graph on n vertices. Efficient algorithm means polynomial algorithms n3, n2, n log n, n New algorithmic paradigm Input: access to a (huge) graph (e.g., for a vertex v, find its neighbors) Bounded number of access. A traditional algorithm Input: a given graph on n vertices. Efficient algorithm means polynomial algorithms n3, n2, n log n, n New algorithmic paradigm Input: access to a (huge) graph (e.g., for a vertex v, find its neighbors) Bounded number of access. The definition of PageRank given by Brin and Page is based on random walks. Random walks in a graph. G : a graph P : transition probability matrix 1 if u v, du := the degree of u. P(u, v) du 0otherwise. A lazy walk: IP W 2 Original definition of PageRank A (bored) surfer • either surf a random webpage with probability α • or surf a linked webpage with probability 1- α α : the jumping constant p ( 1n , 1n ,...., 1n ) (1 ) pW Definition of personalized PageRank Two equivalent ways to define PageRank pr(α,s) (1) p s (1 ) pW s: the seed as a row vector α : the jumping constant s Definition of PageRank Two equivalent ways to define PageRank p=pr(α,s) (1) (2) p s (1 ) pW p (1 )t ( sW t ) t 0 1 1 1 ( , ,...., s= n n n) the (original) PageRank s = some “seed”, e.g., (1, 0,...., 0) personalized PageRank How good is PageRank as a measure of correlationship? Depends on the applications? How “good” is the cut? Isoperimetric properties Isoperimetric properties “What is the shortest curve enclosing a unit area?” In a graph G and an integer m, what is the minimum cut disconnecting a subgraph of ≥ m vertices? In a graph G, what is the minimum cut e(S,V-S) so that e(S,V-S) _____ Vol S is the smallest? How “good” is the cut? Two types of cuts: • Vertex S E(S,V-S) cut • edge cut e(S,V-S) _____ Vol S e(S,V-S) _____ |S| Vol S = Σ deg(v) vεS |S| = Σ 1 vεS V-S S The Cheeger constant for graphs The Cheeger constant e( S , S ) hGhG min S min(vol S , vol S ) The volume of S is vol ( S ) d x xS hGG and its variations are sometimes called “conductance”, “isoperimetric number”, … The Cheeger inequality The Cheeger constant e( S , S ) hG min h( S ) min S S min(vol S , vol S ) The Cheeger inequality hG 2 2hG 2 : the first nontrivial eigenvalue of the xx(normalized) Laplacian of a connected graph. The spectrum of a graph •Adjacency matrix Many ways to define the spectrum of a graph. How are the eigenvalues related to properties of graphs? The spectrum of a graph •Adjacency matrix •Combinatorial Laplacian L D A adjacency matrix diagonal degree matrix Gustav Robert Kirchhoff 1824-1887 The spectrum of a graph •Adjacency matrix •Combinatorial Laplacian adjacency matrix L D A diagonal degree matrix Matrix tree theorem #spanning . trees i i 0 Gustav Robert Kirchhoff 1824-1887 The spectrum of a graph •Adjacency matrix •Combinatorial Laplacian L D A adjacency matrix diagonal degree matrix •Normalized Laplacian Random walks Rate of convergence Gustav Robert Kirchhoff 1824-1887 The spectrum of a graph loopless, simple Discrete Laplace operator 1 f ( x) dx L ( x, y ) { not symmetric in general ( f ( x) f ( y)) y x 1 if x y 1 if x y and x dx •Normalized Laplacian symmetric normalized L( x, y ) with eigenvalues { y 1 if x y 1 if x y and x dxd y 0 0 1 n1 2 y The spectrum of a graph Discrete Laplace operator 1 f ( x) dx L ( x, y ) { not symmetric in general ( f ( x) f ( y)) y x 1 if x y 1 if x y and x dx •Normalized Laplacian symmetric normalized L( x, y ) with eigenvalues { y 1 if x y 1 if x y and x dxd y 0 0 1 n1 2 y dictates many properties of a graph. • expander • diameter • discrepancy • subgraph containment • …. Spectral implications for finding good cuts? Finding a cut by a sweep For f : V (G ) R, order the vertices f (vn ) f (v1 ) f (v2 ) . d v1 d v2 d vn Consider sets S j {v1 , , v j }, j 1,..., n, f and the Cheeger constant of S j . f Define h( f ) min h( S j ) f j Finding a cut by a sweep Using a sweep by the eigenvector, can reduce the exponential number of choices of subsets to a linear number. Finding a cut by a sweep Using a sweep by the eigenvector, can reduce the exponential number of choices of subsets to a linear number. Still, there is a lower bound guarantee by using the Cheeger inequality. h2 2h 2 Four types of Cheeger inequalities. Four proofs using: eigenvectors random walks PageRank heat kernel Leading to four different one-sweep partitioning algorithms. Four proofs of Cheeger inequalities • graph spectral method spectral partition algorithm • random walks • PageRank • heat kernel local partition algorithms Graph partitioning Local graph partitioning What is a local graph partitioning algorithm? A local graph partitioning algorithm finds a small cut near the given seed(s) with running time depending only on the size of the output. Examples of local partitioning Examples of local partitioning Examples of local partitioning Examples of local partitioning Examples of local partitioning h h h Four proofs of Cheeger inequalities • graph spectral method spectral partition algorithm • random walks • PageRank • heat kernel local partition algorithms Four proofs of Cheeger inequalities • graph spectral method Cheeger 60’s, Fiedler ’73 • random walks Alon 86’, Jerrum+Sinclair 89’ Lovasz, Simonovits, 90, 93 Spielman, Teng, 04 • PageRank Andersen, Chung, Lang, 06 • heat kernel Chung, PNAS , 08. The Cheeger inequality Partition algorithm Using eigenvector f , the Cheeger inequality can be stated as 2 h2 2h 2 2 where is the first non-trivial eigenvalue of the Laplacian and is the minimum Cheeger ratio in a sweep using the eigenvector f . Proof of the Cheeger inequality: from definition by Cauchy-Schwarz ineq. from the definition. summation by parts. A Cheeger inequality using random walks Lovász, Simonovits, 90, 93 vol ( S ) W (u , S ) ( S ) 1 du 8 2 k k k Leads to a Cheeger inequality: G 2 h2 2h 8 8 where G is the minimum Cheeger ratio over sweeps by using a lazy walk of k steps from every vertex for an appropriate range of k . A Cheeger inequality using PageRank Using the PageRank vector. Recall the definition of PageRank p=pr(α,s): (1) (2) p s (1 ) pW p (1 ) ( sW ) t t t 0 Organize the random walks by a scalar α. Random walks versus How fast is the convergence to the stationary distribution? For what k, can one have f Wk ? PageRank Choose α to satisfy the required property. A Cheeger inequality using PageRank Using the PageRank vector with seed as a subset S and vol ( S ) vol (G ) / 4, a Cheeger inequality can be obtained : S2 hS 2 hS S 8log s 8log s where S is the Dirichlet eigenvalue of the Laplacian, and S is the minimum Cheeger ratio over sweeps by using the appropriate personalized PageRank with seeds S. Dirichlet eigenvalues for a subset S V S inf 2 ( f ( u ) f ( v )) u v f f ( w) 2 d w w V S over all f satisfying the Dirichlet boundary condition: f (v ) 0 for all v S. Local Cheeger constant for a subset S V V hS min h(T ) T S S A Cheeger inequality using PageRank Using the PageRank vector with seed as a subset S and vol ( S ) vol (G ) / 4, a Cheeger inequality can be obtained : S2 hS 2 hS S 8log s 8log s where S is the Dirichlet eigenvalue of the Laplacian, and S is the minimum Cheeger ratio over sweeps by using personalized PageRank with seed S. Algorithmic aspects of PageRank • Fast approximation algorithm for x personalized PageRank greedy type algorithm, almost linear complexity • Can use the jumping constant to approximate PageRank with a support of the desired size. • Errors can be effectively bounded. A graph partition algorithm using PageRank Given a set S with 1 hS , vol ( S ) vol (G ), 2 randomly choose a vertex v in S. 1 With probability at least , 2 the one-sweep algorithm using f pr ( , v) has an initial segment with the Cheeger constant at most h f O( log | S |). Graph partitioning using PageRank vector. 198,430 nodes and 1,133,512 edges Kevin Lang 2007 Four proofs of Cheeger inequalities • graph spectral method Fiedler ’73, Cheeger, 60’s • random walks Alon 86’ Lovasz, Simonovits, 90, 93 Spielman, Teng, 04 • PageRank Andersen, Chung, Lang, 06 • heat kernel Chung, PNAS , 08. PageRank versus p , s (1 ) k ( sW k ) k 0 Geometric sum heat kernel k ( tW ) t ,s et s k! k 0 Exponential sum PageRank versus p , s (1 ) k ( sW k ) k 0 Geometric sum heat kernel k ( tW ) t ,s et s k! k 0 Exponential sum p (1 ) pW (I W ) t recurrence Heat equation A Cheeger inequality using the heat kernel Theorem: vol (S ) tt2,u / 4 t , u ( S ) ( S ) e du where t ,u is the minimum Cheeger ratio over sweeps by using heat kernel pagerank over all u in S. Theorem: For vol (S ) vol (G)2/ 3 , etS t , S ( S ) ( S ) . 2 Definition of heat kernel 2 k t t H t et ( I tW W 2 ... W k ...) 2 k! t ( I W ) e e tL k t2 2 t I tL L ... (1)k Lk ... 2 k! H t ( I W ) H t t t , s sH t A Cheeger inequality using the heat kernel Using the upper and lower bounds, a Cheeger inequality can be obtained : S 2 hS 2 hS S 8 8 where S is the Dirichlet eigenvalue of the Laplacian, and S is the minimum Cheeger ratio over sweeps by using heat kernel with seeds S for appropriate t. Many applications of PageRank for problems in Graph Theory • Graph drawing using PageRank • Graph embedding using PageRank • Pebbing and routing using PageRank • Covering and packing using PageRank • Relating graph invariants of subgraphs to the host graph using PageRank • Your favorite old problem using PageRank? New Directions in Graph Theory for information networks Topics: • Random graphs with general degrees • pageranks • Algorithmic game theory, graphical games Using: • Spectral methods • Probabilistic methods • Quasirandom …