Mining of Massive Datasets Jure Leskovec, AnandRajaraman, Jeff Ullman Stanford University

Transcription

1 Note to other teachers and users of these slides: We would be delighted if you found this our material useful in giving your own lectures. Feel free to use these slides verbatim, or to modify them to fit your own needs. If you make use of a significant portion of these slides in your own lecture, please include this message, or a link to our web site: Mining of Massive Datasets Jure Leskovec, AnandRajaraman, Jeff Ullman Stanford University

2 Can we identify node groups? (communities, modules, clusters) Nodes: Football Teams Edges: Games played 2

3 NCAA conferences Nodes: Football Teams Edges: Games played 3

4 Can we identify functional modules? Nodes: Proteins Edges: Physical interactions 4

5 Functional modules Nodes: Proteins Edges: Physical interactions 5

6 Can we identify social communities? Nodes: Facebook Users Edges: Friendships 6

7 Social communities High school Summer internship Stanford (Squash) Stanford (Basketball) Nodes: Facebook Users Edges: Friendships 7

8 Non-overlapping vs. overlapping communities 8

9 Nodes Nodes Network Adjacency matrix 9

10 What is the structure of community overlaps: Edge density in the overlaps is higher! Communities as tiles 10

11 Communities in a network This is what we want! 11

12 1) Given a model, we generate the network: Generative model for networks A C B D E F G 2) Given a network, find the best model H A C B D E H F G Generative model for networks 12

13 Goal:Define a model that can generate networks The model will have a set of parameters that we will later want to estimate (and detect communities) Generative model for networks A C B D E F G Q: Given a set of nodes, how do communities generate edges of the network? H 13

14 Communities, C p A p B Model Memberships, M Nodes, V Model Generative model B(V, C, M, {p c }) for graphs: Nodes V, Communities C, Memberships M Network Each community chas a single probability p c Later we fit the model to networks to detect communities 14

15 Communities, C Memberships, M p A p B Model Nodes, V Community Affiliations AGM generates the links: For each For each pair of nodes in community, we connect them with prob. The overall edge probability is: P ( u, v) = 1 (1 ) p c c M u M v If, share nocommunities:, Network Think of this as an OR function: If at least 1 community says YES we create an edge set of communities node belongs to 15

16 Model Network 16

17 AGM can express a variety of community structures: Non-overlapping, Overlapping, Nested 17

18

19 Detecting communities with AGM: A C B D E F G H Given a Graph, find the Model 1) Affiliation graph M 2) Number of communities C 3) Parameters p c 19

20 Maximum Likelihood Principle (MLE): Given:Data Assumption:Data is generated by some model model model parameters Want to estimate : The probability that our model (with parameters ) generated the data Now let s find the most likely model that could have generated the data: arg max 20

21 Imagine we are given a set of coin flips Task:Figure out the bias of a coin! Data:Sequence of coin flips:,,,,,,, Model: return 1 with prob. Θ,else return 0 What is? Assuming coin flips are independent So, What is? Simple, Then, For example:... What did we learn?our data was most likely generated by coin with bias / / 21

22 How do we do MLE for graphs? Model generates a probabilistic adjacency matrix We then flip all the entries of the probabilistic matrix to obtain the binary adjacency matrix For every pair of nodes, AGM gives the prob. of them being linked Flip biased coins The likelihood of AGM generating graph G: P( G Θ) = Π ( u, v) E P( u, v) Π ( u, v) E (1 P( u, v)) 22

23 Given graph G(V,E)and Θ,we calculate likelihood that Θ generated G: P(G Θ) G A B Θ=B(V, C, M, {p c }) P(G Θ) G P( G Θ) = Π ( u, v) E P( u, v) Π ( u, v) E (1 P( u, v)) 23

24 Our goal:find,,, such that: arg max Θ P( ) AGM Θ How do we find,,, that maximizes the likelihood? 24

25 Our goal is to find,,, such that: arg max,,,,,, Problem: Finding Bmeans finding the bipartite affiliation network. There is no nice way to do this. Fitting,,, is too hard, let s change the model (so it is easier to fit)! 25

26 Relaxation: Memberships have strengths u v :The membership strength of node to community ( : no membership) Each community links nodes independently:, 26

27 j Community membership strength matrix Nodes Communities, Probability of connection is proportional to the product of strengths Notice: If one node doesn t belong to the community ( 0) then, Prob. that at least one common community links the nodes:,, strength of s membership to vector of community membership strengths of 27

28 : : : Community links nodes, independently:, Then prob. at least one common links them:,, Example matrix: Then:. And:,.. But:,., Node community membership strengths 28

29 Task: Given a network,, estimate Find that maximizes the likelihood:,,,, where:, Many times we take the logarithm of the likelihood, and call it log-likelihood: Goal: Find that maximizes : 29

30 Compute gradient of a single row of : Coordinate gradient ascent: Iterate over the rows of : Compute gradient of row (while keeping others fixed) Update the row :.. Set out outgoing neighbors Project back to a non-negative vector: If : This is slow!computing takes linear time! 30

31 However, we notice: We cache So, computing in the degree of now takes linear time In networks degree of a node is much smaller to the total number of nodes in the network, so this is a significant speedup! 31

32 Time (Sec.) Link Clustering Clique Percolation MMSB BigCLAM Parallel BigCLAM Number of nodes ( 10 3 ) BigCLAMtakes 5 minutes for 300k node nets Other methods take 10 days Can process networks with 100M edges! 32

33

34 34

35 Extension: Make community membership edges directed! Outgoing membership: Nodes sends edges Incoming membership: Node receives edges 35

36 36

37 Everything is almost the same except now we have 2 matrices: and out-going community memberships in-coming community memberships Edge prob.:, Network log-likelihood: which we optimize the same way as before 37

38 38

39 Overlapping Community Detection at Scale: A Nonnegative Matrix Factorization Approachby J. Yang, J. Leskovec.ACM International Conference on Web Search and Data Mining (WSDM), Detecting Cohesive and 2-mode Communities in Directed and Undirected Networksby J. Yang, J. McAuley, J. Leskovec.ACM International Conference on Web Search and Data Mining (WSDM), Community Detection in Networks with Node Attributesby J. Yang, J. McAuley, J. Leskovec.IEEE International Conference On Data Mining (ICDM),