CS224W: Analysis of Networks Jure Leskovec, Stanford University

Transcription

1 CS224W: Analysis of Networks Jure Leskovec, Stanford University

2 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 2 Observations Models Algorithms Small diameter, Edge clustering Patterns of signed edge creation Viral Marketing, Blogosphere, Memetracking Scale-Free Densification power law, Shrinking diameters Strength of weak ties, Core-periphery Erdös-Renyi model, Small-world model Structural balance, Theory of status Independent cascade model, Game theoretic model Preferential attachment, Copying model Microscopic model of evolving networks Kronecker Graphs Decentralized search Models for predicting edge signs Influence maximization, Outbreak detection, LIM PageRank, Hubs and authorities Link prediction, Supervised random walks Community detection: Girvan-Newman, Modularity

3 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 3 We often think of networks looking like this: What led to such a conceptual picture?

4 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 4 How does information flow through the network? What structurally distinct roles do nodes play? What roles do different links (short vs. long) play? How do people find out about new jobs? Mark Granovetter, part of his PhD in 1960s People find the information through personal contacts But: Contacts were often acquaintances rather than close friends This is surprising: One would expect your friends to help you out more than casual acquaintances Why is it that acquaintances are most helpful?

5 [Granovetter 73] 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 5 Two perspectives on friendships: Structural: Friendships span different parts of the network Interpersonal: Friendship between two people is either strong or weak Structural role: Triadic Closure a b Which edge is more likely, a-b or a-c? c If two people in a network have a friend in common, then there is an increased likelihood they will become friends themselves.

6 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 6 Granovetter makes a connection between social and structural role of an edge First point: Structure Structurally embedded edges are also socially strong Long-range edges spanning different parts of the network are socially weak Second point: Information Long-range edges allow you to gather information from different parts of the network and get a job Structurally embedded edges are heavily redundant in terms of information access S S S a Weak W b S Strong

7 Triadic closure == High clustering coefficient Reasons for triadic closure: If B and C have a friend A in common, then: B is more likely to meet C (since they both spend time with A) B and C trust each other (since they have a friend in common) A has incentive to bring B and C together (as it is hard for A to maintain two disjoint relationships) Empirical study by Bearman and Moody: Teenage girls with low clustering coefficient are more likely to contemplate suicide 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 7 B A C

8 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 8 Define: Bridge edge If removed, it disconnects the graph Define: Local bridge Edge of Span > 2 (Span of an edge is the distance of the edge endpoints if the edge is deleted. Local bridges with long span are like real bridges) Define: Two types of edges: Strong (friend), Weak (acquaintance) Define: Strong triadic closure: Two strong ties imply a third edge Fact: If strong triadic closure is satisfied then local bridges are weak ties! S Bridge a a b Local bridge S S a S W W b Edge: W or S b S S S

9 Claim: If node A satisfies Strong Triadic Closure and is involved in at least two strong ties, then any local bridge adjacent to A must be a weak tie. Proof by contradiction: Assume A satisfies Strong Triadic Closure and has 2 strong ties Let A B be local bridge and a strong tie Then B C must exist because of Strong Triadic Closure But then A B is not a bridge! (since B-C must be connected due to Strong Triadic Closure property) 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 9 C S S S A A S S B

10 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 10 For many years Granovetter s theory was not tested But, today we have large who-talks-to-whom graphs: , Messenger, Cell phones, Facebook Onnela et al. 2007: Cell-phone network of 20% of country s population Edge strength: # phone calls

11 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 11 Edge overlap: O &' = N(i) N j N(i) N j N(i) a set of neighbors of node i Overlap = 0 when an edge is a local bridge

12 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 12 Cell-phone network Observation: Highly used links have high overlap! Legend: True: The data Permuted strengths: Keep the network structure but randomly reassign edge strengths Neighborhood overlap Permuted strengths True Edge strength (#calls)

13 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 13 Real edge strengths in mobile call graph Strong ties are more embedded (have higher overlap)

14 Same network, same set of edge strengths but now strengths are randomly shuffled 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 14

15 Size of largest component Low disconnects the network sooner Removing links by strength (#calls) Low to high High to low Fraction of removed links Conceptual picture of network structure 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 15

16 Size of largest component Low disconnects the network sooner Removing links based on overlap Low to high High to low Fraction of removed links Conceptual picture of network structure 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 16

17 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 17 Granovetter s theory leads to the following conceptual picture of networks Strong ties Weak ties

18

19 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 19 Granovetter s theory suggest that networks are composed of tightly connected sets of nodes Network communities: Communities, clusters, groups, modules Sets of nodes with lots of connections inside and few to outside (the rest of the network)

20 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 20 How to automatically find such densely connected groups of nodes? Ideally such automatically detected clusters would then correspond to real groups For example: Communities, clusters, groups, modules

21 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 21 Zachary s Karate club network: Observe social ties and rivalries in a university karate club During his observation, conflicts led the group to split Split could be explained by a minimum cut in the network

22 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 22 Find micro-markets by partitioning the query-to-advertiser graph in web search: query advertiser

23 Can we identify node groups? (communities, modules, clusters) 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, Nodes: Teams Edges: Games played 23

24 11/13/17 Nodes: Teams Edges: Games played Jure Leskovec, Stanford CS224W: Analysis of Networks, 24 NCAA conferences

25 Can we identify social communities? 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, Nodes: Users Edges: Friendships 25

26 High school Company Stanford (Squash) Stanford (Basketball) Social communities 11/13/17 Nodes: Users Edges: Friendships Jure Leskovec, Stanford CS224W: Analysis of Networks, 26

27 Can we identify functional modules? 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, Nodes: Proteins Edges: Interactions 27

28 11/13/17 Nodes: Proteins Edges: Interactions Jure Leskovec, Stanford CS224W: Analysis of Networks, 28 Functional modules

29 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 29 How to find communities? We will work with undirected (unweighted) networks

30 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 30 Edge betweenness: Number of shortest paths passing over the edge Intuition: b=16 b=7.5 Edge strengths (call volume) in a real network Edge betweenness in a real network

31 [Girvan-Newman 02] 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 31 Divisive hierarchical clustering based on the notion of edge betweenness: Number of shortest paths passing through the edge Girvan-Newman Algorithm: Undirected unweighted networks Repeat until no edges are left: Calculate betweennessof edges Remove edges with highest betweenness Connected components are communities Gives a hierarchical decomposition of the network

32 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, Need to re-compute betweenness at every step

33 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 33 Step 1: Step 2: Step 3: Hierarchical network decomposition:

34 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, How to compute betweenness? 2. How to select the number of clusters?

35 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 35 Want to compute betweenness of paths starting at node A Breadth first search starting from A:

36 Forward step: Count the number of shortest paths from A to all other nodes of the network 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 36

37 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 37 Backward step: Compute betweenness: If there are multiple paths count them fractionally The algorithm: Add edge flows: -- node flow = 1+ child edges -- split the flow up based on the parent value Repeat the BFS procedure for each starting node U 1 path to K. Split evenly paths to J Split 1:2 1+1 paths to H Split evenly

38 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 38 Backward step: Compute betweenness: If there are multiple paths count them fractionally The algorithm: Add edge flows: -- node flow = 1+ child edges -- split the flow up based on the parent value Repeat the BFS procedure for each starting node U 1 path to K. Split evenly paths to J Split 1:2 1+1 paths to H Split evenly

39 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, How to compute betweenness? 2. How to select the number of clusters?

40 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 40 Communities: sets of tightly connected nodes Define: Modularity Q A measure of how well a network is partitioned into communities Given a partitioning of the network into groups s S: Q µ sî S [ (# edges within group s) (expected # edges within group s) ] Need a null model!

41 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 41 Given real G on n nodes and m edges, construct rewired network G Same degree distribution but random connections Consider G as a multigraph The expected number of edges between nodes i and j of degrees k i and k j equals to: k i k j = k ik j 2m 2m The expected number of edges in (multigraph) G : = 1 k i k j 2 i N j N = 1 1 k 2m 2 2m i j N k j = 1 2m 2m = m 4m i N = i j Note: F k H H I = 2m

42 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 42 Modularity of partitioning S of graph G: Q µ sî S [ (# edges within group s) (expected # edges within group s) ] Q G, S = 1 A 2m ij k ik j s S i s j s 2m Normalizing cost.: -1<Q<1 A ij = 1 if i j, 0 else Modularity values take range [ 1,1] It is positive if the number of edges within groups exceeds the expected number <Q means significant community structure

43 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 43 Modularity is useful for selecting the number of clusters: Q Why not optimize Modularity directly?

44

45 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 45 Let s split the graph into 2 communities! Want to directly optimize modularity! arg max R Q G, S = V A WX &' Z [Z \ ] R & ] ' ] WX Community membership vector s: s i = 1 if node i is in community 1-1 if node i is in community -1 Q G, s = V A WX &' Z [Z \ & I ' I WX = V A `X &' Z [Z \ &,' I s WX &s ' s & s ' ] [ ] \ _V W = 1.. if s i=s j 0.. else

46 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 46 Define: Modularity matrix: B ij = A ij k ik j 2m Membership: s = { 1, +1} Then: Q G, s = V A `X &' Z [Z \ & I ' I WX = V B `X &,' I &' s & s ' = V s `X & & ' B &' s ' = V `X sf Bs = B i s Note: each row/col of B sums to 0: A ij = k i j, k ik j j = k 2m i k j j Task: Find sî{-1,+1} n that maximizes Q(G,s) 2m = k i s &s '

47 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 47 Symmetric matrix A That is positive semi-definite: A = U U T Then solutions λ, x to equation A x = λ x : Eigenvectors x i ordered by the magnitude of their corresponding eigenvalues λ & (λ V λ W λ n ) x i are orthonormal (orthogonal and unit length) In other words, x i form a coordinate system (basis) If A is positive-semidefinite: λ & 0 (and they always exist) Eigen Decomposition theorem: Can rewrite matrix A in terms of its eigenvectors and eigenvalues: A = T i x i λ & x i

48 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 48 Rewrite: Q G, s = V `X sq Bs in terms of its eigenvectors and eigenvalues: n = s q F x & λ & x & f &tv n s = F s f x & λ & x & f s &tv n = F s f x & W λ & &tv So, if there would be no other constraints on s then to maximize Q, we make s = x n x Why? Because λ n λ nu1 2 Remember s has fixed euclidean length! Assigns all weight in the sum to λ n (largest eigenvalue) All other s T x i terms are zero because of orthonormality s x 1

49 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 49 Let s consider only the first term in the summation (because λ n is the largest): n arg max Q G, s = &tv s f x W & λ & s f x W n λ n ] Let s maximize: s T x n where s j Î{-1,+1} To do this, we set: s j = x +1 1 if x n,j 0 (j th coordinate of x n 0) if x n,j < 0 (j th coordinate of x n < 0) Continue the bisection hierarchically

50 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 50 Fast Modularity Optimization Algorithm: Find leading eigenvector x n of modularity matrix B Divide the nodes by the signs of the elements of x n Repeat hierarchically until: If a proposed split does not cause modularity to increase, declare community indivisible and do not split it If all communities are indivisible, stop How to find x n? Power method! Start with random v (0), repeat : When converged (v (t) v (t+1) ), set x n = v (t) ( t+ 1) v = Bv Bv ( t) ( t)

51 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 51 Girvan-Newman: Based on the strength of weak ties Remove edge of highest betweenness Modularity: Overall quality of the partitioning of a graph Use to determine the number of communities Fast modularity optimization: Transform the modularity optimization to a eigenvalue problem

52

53 [Ron Burt] 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 53 Who is better off, Robert or James?

54 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 54 Few structural holes Many structural holes Structural Holes provide ego with access to novel information, power, freedom

55 The network constraint measure [Burt]: c To what extent are person s contacts redundant Low: disconnected contacts High: contacts that are close or strongly tied = å i c p ik = å k p ij é ê p ë + å( p ) ik pkj i ij ij j j k p uv prop. of u s energy invested in relationship with v j p uv = 1/d u 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 55 ú û ù 2 2 p 12 =¼ 2 i 1 4 k p 15 =¼ p 25 =½ p uv j

56 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 56 Network constraint: James: c J = Robert: c R = Constraint: To what extent are person s contacts redundant Low: disconnected contacts High: contacts that are close or strongly tied

57 [Ron Burt] 11/13/17 Jure Leskovec, Stanford CS224W: Analysis of Networks, 57