information networks power laws and network models lecture 3

24
Information Networks Power Laws and Network Models Lecture 3

Upload: lindsey-bennett

Post on 25-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Information Networks Power Laws and Network Models Lecture 3

Information Networks

Power Laws and Network Models

Lecture 3

Page 2: Information Networks Power Laws and Network Models Lecture 3

Erdös-Renyi Random Graphs

For each pair of vertices (i,j), generate the edge (i,j) independently with probability p degree sequence: Binomial

in the limit: Poisson

real life networks: power-law distribution

Low clustering coefficient: C = z/n Short paths, but not easy to navigate Too random…

knk p1pk

np)k;B(n; p(k)

zk

ek!z

z)P(k;p(k)

p(k) = Ck-α

Page 3: Information Networks Power Laws and Network Models Lecture 3

Departing from the ER model

We need models that better capture the characteristics of real graphs degree sequences clustering coefficient short paths

Page 4: Information Networks Power Laws and Network Models Lecture 3

Graphs with given degree sequences

Assume that we are given the degree sequence [d1,d2,…,dn]

Create di copies of node i

Take a random matching (pairing) of the copies self-loops and multiple edges are allowed

Uniform distribution over the graphs with the given degree sequence

Page 5: Information Networks Power Laws and Network Models Lecture 3

The giant component

The phase transition for this model happens when

The clustering coefficient is given by

0k

k 02)pk(k

2

2

2

d

dd

nz

C

pk: fraction of nodes with degree k

Page 6: Information Networks Power Laws and Network Models Lecture 3

Power-law graphs

The critical value for the exponent α is

The clustering coefficient is

When α<7/3 the clustering coefficient increases with n

3.4788...α

βnC 1α73α

β

Page 7: Information Networks Power Laws and Network Models Lecture 3

A different approach

Given a degree sequence [d1,d2,…,dn], where d1≥d2≥…≥dn, there exists a simple graph with this sequence (the sequence is realizable) if and only if

There exists a connected graph with this sequence if and only if

k,dmin1)k(kd i

n

1ki

k

1ii

k

1ii 1)2(nd

Page 8: Information Networks Power Laws and Network Models Lecture 3

Getting a random graph

Perform a random walk on the space of all possible simple connected graphs at each step swap the endpoints of two

randomly picked edges

Page 9: Information Networks Power Laws and Network Models Lecture 3

Graphs with given degree sequence

The problem is that these graphs are too contrived

It would be more interesting if the network structure emerged as a side product of a stochastic process

Page 10: Information Networks Power Laws and Network Models Lecture 3

Power Laws

Two quantities y and x are related by a power law if

A (continuous) random variable X follows a power-law distribution if it has density function

Cumulative function

αxy

αCxf(x)

1)(αx1α

CxXP

Page 11: Information Networks Power Laws and Network Models Lecture 3

Normalization constant

Assuming a minimum value xmin

The density function becomes

1aminx1αC

α

minmin xx

x1)-(α

f(x)

Page 12: Information Networks Power Laws and Network Models Lecture 3

Pareto distribution

Pareto distribution is pretty much the same but we have

and we usually we require

βxC'xXP

minxx

Page 13: Information Networks Power Laws and Network Models Lecture 3

Zipf’s Law

A random variable X follows Zipf’s law if the r-th largest value xr satisfies

Same as requiring a Pareto distribution

γr rx

γ1xxXP

Page 14: Information Networks Power Laws and Network Models Lecture 3

Power laws are ubiquitous

Page 15: Information Networks Power Laws and Network Models Lecture 3

But not everything is power law

Page 16: Information Networks Power Laws and Network Models Lecture 3

Measuring power laws

Simple log-log plot gives poor estimate

Page 17: Information Networks Power Laws and Network Models Lecture 3

Logarithmic binning

Bin the observations in bins of exponential size

Page 18: Information Networks Power Laws and Network Models Lecture 3

Cumulative distribution

Page 19: Information Networks Power Laws and Network Models Lecture 3

Computing the exponent

Least squares fit of a line

Maximum likelihood estimate

1n

1i min

i

xx

lnn1α

Page 20: Information Networks Power Laws and Network Models Lecture 3

Power-law properties

First moment

Second moment

minx2α1α

x

22min

x3α1α

x

Page 21: Information Networks Power Laws and Network Models Lecture 3

Scale Free property

A power law holds in all scale f(bx) = g(b)p(x)

Page 22: Information Networks Power Laws and Network Models Lecture 3

The 80/20 rule

Cumulative distribution is top-heavy

Page 23: Information Networks Power Laws and Network Models Lecture 3

The discrete case

αζk

Ckpα

αk

αk k1αkα,B1αkα,CBp

2α1α

k

3α2α

1αk2

2

Page 24: Information Networks Power Laws and Network Models Lecture 3

References

M. E. J. Newman, The structure and function of complex networks, SIAM Reviews, 45(2): 167-256, 2003

M. E. J. Newman, Power laws, Pareto distributions and Zipf's law, Contemporary Physics

M. Mitzenmacher, A Brief History of Generative Models for Power Law and Lognormal Distributions, Internet Mathematics

M. Mihail, N. Vishnoi ,On Generating Graphs with Prescribed Degree Sequences for Complex Network Modeling Applications, Position Paper, ARACNE (Approx. and Randomized Algorithms for Communication Networks) 2002, Rome, IT, 2002.