+ All Categories
Home > Documents > Flexible Graph Models for Complex Networks

Flexible Graph Models for Complex Networks

Date post: 25-Feb-2016
Category:
Upload: mac
View: 52 times
Download: 0 times
Share this document with a friend
Description:
Joint with Bradley Green 2 , Milena Mihail 3 (advisor). 1. School of Mathematics, Georgia Tech, supported by ARC and ACO. 2. School of Mathematics, Georgia Tech, supported by NSF VIGRE. 3. School of Computer Science, Georgia Tech, supported by NSF CCF 05 & 08. - PowerPoint PPT Presentation
12
Flexible Graph Models for Complex Networks rks: Internet and its applications: WWW, content sharing Other: further socia Yorgos Amanatidis 1 ol of Mathematics, Georgia Tech, supported by ARC and ACO. with Bradley Green 2 , Milena Mihail 3 (advisor) l of Mathematics, Georgia Tech, supported by NSF VIGRE. of Computer Science, Georgia Tech, supported by NSF CCF e Graph Models: Few parameters e tuning result in vastly distinct features bly annotated) network structure and function. al to explain, predict and control network fun ocus on efficiency of model realizability, sin ng rate of the related applications is quite d Poster Outline: Pages 2,3: Brief Summary of Standard Complex Network Models Page 4: Two Low Entropy Cases: Disassortativity and Sparse Cuts Pages 5,6,7,8: The Joint Degree Matrix Realizability Problem (addressing the case of disassortative networks) Pages 9,10,11: Random Kernel (similarity) Graphs (towards generalizing standard models to include distinct special cases such as sparse cuts and disassortativity)
Transcript
Page 1: Flexible Graph Models for Complex Networks

Flexible Graph Models for Complex Networks

Complex Networks: Internet and its applications: WWW, content sharing, social online Other: further social, biological

Yorgos Amanatidis1

1. School of Mathematics, Georgia Tech, supported by ARC and ACO.

Joint with Bradley Green2, Milena Mihail3 (advisor)

2. School of Mathematics, Georgia Tech, supported by NSF VIGRE.3. School of Computer Science, Georgia Tech, supported by NSF CCF 05 & 08.

Flexible Graph Models: Few parameters whose fine tuning result in vastly distinct features of (possibly annotated) network structure and function. Fundamental to explain, predict and control network function. We also focus on efficiency of model realizability, sincethe scaling rate of the related applications is quite dramatic.

Poster Outline:Pages 2,3: Brief Summary of Standard Complex Network ModelsPage 4: Two Low Entropy Cases: Disassortativity and Sparse CutsPages 5,6,7,8: The Joint Degree Matrix Realizability Problem (addressing the case of disassortative networks)Pages 9,10,11: Random Kernel (similarity) Graphs (towards generalizing standard models to include distinct special cases such as sparse cuts and disassortativity) Page 12: Network Evolution and Efficient Implementation

Page 2: Flexible Graph Models for Complex Networks

2

Structure & Function Early ObservationsStructure: Heavy Tailed Statistics

Function: Small World Phenomenon

In a clustered graph of size n,O(n) random links decrease the diameter to O(logn)

AS Level Internet(each node is an ISP)

Gene-Protein Interaction network for the Yeast

Nearly constant average degree (as the network scaled)but no concentration of degrees around their mean.Suggest paradigm shift from Erdős-Rényi models.

Local clustering and small diameter, especially navigability.Suggest function as an integral part of model definition and evaluation.Navigability stresses importance of accurate parametrization.

When the O(n) links are added with then paths of length O(polylogn) can be found with local search,iff r is the dimension of the lattice.

Page 3: Flexible Graph Models for Complex Networks

3

Complex Network Models 1st Generation

Macroscopic (see R. Durrett’s book 08)

Microscopic (see M. Jackson’s book 08)

Random graphs capturing heavy tailed statistics. Further effort to verify the modelsby establishing efficient network function (diameter,conductance, navigability)Advantages: conceptually simple,General, efficient to implement.Disadvantage: missing semantics

Explanatory: evolutionary/optimization/incentive-driven processes.Further effort to verify the models by establishing agreement with observed structural properties.Advantage: fully annotated and explanatory.Disadvantages: lack of generality, particularly inefficient to simulateSuccessful: in cases where network elements have strongly distinct semantics.Example: Internet router connectivity within an ISP. There is vast variance in the capacities of different routers

Successful: when all network elements are roughly similar,Thus the missing semantics are not particularly important.Example: AS Level Internet (all ISP’s are more or less the same).

Page 4: Flexible Graph Models for Complex Networks

4

Existence of Sparse Cuts

Cases where Macroscopic Models Fail

Router connectivity of a singleprovider (picture reflecting real data):low degree nodes correspondto high bandwidth routers and are in the “center” of the network.High degree nodes are in the “periphery”. Is in sharp contrast with random graphs.

Heuristic: in addition tomatching expected degrees,match number of links between nodes in distinct degree classes(picture reflecting syntheticdata following heuristic).

Flickr relations Patent collaborations (Boston)Sparse cuts indicate a subgroup with special semantics. Sparse cuts of small size are of particular importance (law of the few):an emerging trend, a group of terrorists, potential starting of an epidemic.They are present in many complex networks, well beyond what standard random graph models predict.

Negative Assortativity

Page 5: Flexible Graph Models for Complex Networks

5

Graphic Realizations of Joint Degree Matricesaddressing negative assortativity

The Joint-Degree Matrix Realization Problem:

The J-D Matrix Connected Realization Problem:

Page 6: Flexible Graph Models for Complex Networks

6

Theorem 1 [A,Green, Mihail ’08]:(generalization of Erdős-Gallai, Havel-Hakimi)

Theorem 2 [A,Green,Mihail ’08]:

Page 7: Flexible Graph Models for Complex Networks

7

Remarks: Proofs of both theorems are algorithmic. The algorithm for Theorem 1, is greedy. However, it involves augmentations along alternating structures, that are neither simple augmentations, nor straightforward augmenting paths. Key to the algorithm is a certain balanced degree invariant.For Theorem 2, the necessary and sufficient conditions are fairly complex, and of exponential size. We introduce a novel poly-time recursive algorithm that searches for suitable local graph modifications to construct a connected graph,or identifies a necessary condition that fails to hold. The latter is illustrated below.

1V2V

5V

3V

4V

4

2

11

1

1

1

2

Think of these as a single component.

0 4 0 2 14 0 1 0 10 1 0 1 22 0 1 0 11 1 2 1 0

D

2

4

3

1 2 3

4 5

( ) 1, ( ) 2, ( ) 2,( ) 4, ( ) 5

d V d V d Vd V d V

Now we have 9 available edges for 11 vertices, so there is no way to obtain a connected graph even if the “red blob” is connected. This is a typical example of a certificate that no connected graph exists.

Page 8: Flexible Graph Models for Complex Networks

8

Open Problems

Page 9: Flexible Graph Models for Complex Networks

9

Model :

Random Kernel Function Graphs

Model (slightly simplified for this poster):

Motivation: Nodes have semantics expressedby values on d distinct attributes, and are generated from a general distribution .Connections between nodes have semanticsexpressed by a kernel function.

Page 10: Flexible Graph Models for Complex Networks

10

Model with

Known Results for Random Kernel GraphsBollobás, Janson, Riordan 07:

Semiclosed characterization of connectivity andphase transition in terms of (under mild assumptions) .

Subsumes most known sparse random graph modelsby suitable definitions of .Main point: vast classes of sparse random graphsreduce locally (and can be analyzed) like Erdős-Rényi.

Young, Mihail 08:

Characterization (with concentration) of degree distribution, diameter, clustering, in terms of .Includes heavy tailed models and small world phenomenonby suitable definition of (without assumptions).Semiclosed characterization of conductance, in terms of (conjectured if a.e.)

Main point: there is a random graph model with clear semantics,conceptually simple and efficient to implement,that captures many critical structural and functional aspects of complex networks, and generates quite distinct classes graphs.

Main challenge:

whose characterization reduces to

Page 11: Flexible Graph Models for Complex Networks

11

Probability Distributions & Kernel Functions towards modeling Sparse Cuts

Probability Distributions & Kernel Functions towards modeling Negative Assortativity

In theory, for wide ranges of network characteristics(using Young & Mihail 08), and experimentally A & Mihail 08).

Experimentally A & Mihail 08.

Page 12: Flexible Graph Models for Complex Networks

12

The general allows each instance to evolve naturally from previous instances.

Evolution

Efficient Implementation

In general, realizing an instance of requires experiments.For where expected degrees can be computedefficiently (including inner product graphs), this can be reduced to by “approximating” the distribution as in the configuration model.It is interesting to prove guarantees for such approximations. In principle, the introduced dependencies only force degree concentration, and we expect good behavior.


Recommended