Data mining --- mining graphs

Transcription

1 Data mining --- mining graphs University of South Florida Xiaoning Qian

2 Today s Lecture 1. Complex networks 2. Graph representation for networks 3. Markov chain 4. Viral propagation 5. Google s PageRank

3 New Science of Networks 1. M Faloutsos, P Faloutsos, C Faloutsos, On power-law relationships of the Internet topology. Comput. Commun. Rev. 29: , JM Kleinberg, Navigation in a small world it is easier to find short chains between points in some networks than others. Nature, 406:845, AL Barabasi, Linked: The New Science of Networks. Cambridge, MA, DJ Watts, The new science of networks. Annu. Rev. Sociol. 30:243 70, DL Alderson, Catching the network science bug: Insight and opportunity for the operations researcher. Operations Research 56(5): , MEJ Newman, Networks: An Introduction. Oxford, 2010

4 Network Science The first network/graph problem Find a tour crossing every bridge just once, Euler, 1735 Bridges of Königsberg

5 Network Science 1. S Milgram, The small world problem. Psychol. Today 2:60 67, 1967 New Science Unprecedented number of empirical networks Much larger scale networks Visualization does not convey enough information Computer are much more powerful Highly interdisciplinary

6 Network Science Mining networks/graphs Topology/structure of complex networks Global: degrees, centrality, connectivity, etc. Scale-free (power-law) networks: 6 degree separation? Local: clustering (community), network motifs, etc. Dynamics/behavior of complex networks Global: the topological effect on dynamics How information, virus, disease, rumors, etc. propagate? Local: how individual nodes behave

7 Complex Networks (Yeast PPI)

8 Complex Networks (Yeast signaling)

9 Complex Networks (food web)

10 Complex Networks (friendship)

11 Complex Networks (romantic relation)

12 Complex Networks (author citation)

13 Complex Networks (Internet)

14 Complex Networks (Web)

15 Mathematics of Networks (Graphs) What is a network/graph? A collection of vertices/nodes joined by edges Different types of vertices and edges: Directed vs. Undirected Weighed vs. Binary Labeled vs. Nonlabeled Bipartite graphs Hypergraphs Mathematically, G = {V, E}

16 Mathematics of Networks (Graphs) Undirected network: <v i, v j > E => < v j, v i > E

17 Mathematics of Networks (Graphs) Adjacency matrix L Symmetric for undirected graphs Square matrix for (self-)graphs; rectangular for bipartite graphs L ij = e ij if <v i, v j > E Matrix analysis for graph mining! Simple graphs, connected graphs, complete graphs,

18 Mathematics of Networks (Graphs) Node degree c i The number of edges incident with vertex v i Neighbor set Input-, output-degrees Degree distributions (power-law) Trail (distinct edges), path (distinct nodes), cycle, cut,

19 Markov chain Sequential data

20 What is a Markov chain? Finite Markov chain -- (Q, P) Q = {q 1, q 2,, q s } : a finite set of states P : state transition probability matrix Given a sequence of observations: The probability of the sequence is: For first-order time-homogeneous Markov chain: Hence,

21 What is a Markov chain? Finite Markov chain -- (Q, P) Q = {B, q 1, q 2,, q s } : a finite set of states P : state transition probability matrix initial state probability: The probability of a sequence can be expressed with P: Note: The output are states at each time -- states are observable!!

22 An example 3-state Markov chain model for the weather: R 0.1 S C Q = {Rain (or snow), Cloudy, Sunny}; P is given in the figure; Initial state probability 0.6

23 Chapman-Kolmogorov Equation Chapman-Kolmogorov equation p(x n ) = P (n-1) p(x 1 ) Limiting distribution (stationary/steady-state distribution) Irreducibility, Periodicity, Ergocity p = P p How to solve p? Eigen-decomposition of P Power method

24 Random walk on graphs Random walk on graphs (network diffusion) is a Markov process.

25 What s behind Google? The algorithm of Google---PageRank

26 PageRank What is an important Webpage? There are many Webpages pointing to it Important Webpage point to more important Webpage Importance diffuses based on links between Webpages Vertices: Webpages; Edges: hyperlinks; HITS: JM Kleinberg

27 PageRank Diffusion (Random walk) on Web λ λ p i : importance for page i; L ij : link from page j to i; Hence, the problem becomes a Markov chain problem (diffusion process): λ λ

28 PageRank Diffusion (Random walk with restart) on Web λ λ p i : importance for page i; L ij : link from page j to i;

29 PageRank Diffusion (Random walk with restart) on Web Pseudocount λ λ Diffusion factor Matrix form: λ λ

30 PageRank How do we solve this? λ λ Note that p is simply for ranking and the absolute values are not critical! WLOG, we assume Hence, the problem becomes a Markov chain problem (diffusion process): λ λ

31 Viral propagation How does the virus spread over the network? Will it become an epidemic outbreak? How fast the virus will die out or become epidemic? How we should design robust networks to prevent cascading failures? * D Chakrabati, Tools for large graph mining. Ph.D. Thesis, CMU, 2005

32 Mathematical Epidemiology SIR (Susceptible-Infective-Recovered) model SIS (Susceptible-Infective-Susceptible) model Catching the disease from Infective neighbors (birth rate): β Recover rate: δ Epidemic threshold: τ

33 SIS model SIS model is again a Markov process!?? Sum and Product rules in probability!!

34 SIS model Sum and Product rules in probability!!

35 SIS model With appropriate approximations, we can derive p(v i t =susceptible) = p(v i t-1 =susceptible) ζ i + p(v i t-1 =infective) δ 1-p(v i t ) = [1-p(v i t-1 )] ζ i + p(v i t-1 ) δ and p t =[ βl + (1-δ)I ] p t-1 Sum and Product rules in probability!!

36 SIS model With appropriate approximations, we can derive p t =[ βl + (1-δ)I ] p t-1 Eigen-decomposition of the matrix S= [ βl + (1-δ)I ] p p 0 L Hence, L

37 SIS model With appropriate approximations, we can derive p t =[ βl + (1-δ)I ] p t-1 Eigen-decomposition of the matrix S= [ βl + (1-δ)I ] Hence, L Epidemic threshold: L

38 Summary Networks/graphs are everywhere and require new tools to study them efficiently and effectively. Random walk (Markov chain) on graphs and its extension can be a useful technique to mine complex networks/graphs PageRank Viral propagation Have you learned anything? :) I am teaching Biological Network Analysis, Spring 2012.