0% found this document useful (0 votes)

45 views93 pages

Network Dynamics 2013

This document provides an overview of the course "Network Dynamics" being taught by Giacomo Como. The course will consist of 18 two-hour meetings over 9 weeks, covering topics like network structure, dynamics on networks, random networks, and complex network properties. It will include lectures, exercises, homework assignments, and a final student presentation. The course will use reference materials on network science and dynamics, and introduce mathematical representations of network structure.

Uploaded by

Banyan Vines

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

45 views93 pages

Network Dynamics 2013

Uploaded by

Banyan Vines

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 93

Network Dynamics:

an Overview
Giacomo Como
Department of Automatic Control, LTH
www.control.lth.se/Staff/GiacomoComo/
[email protected]

October 1, 2013

Network Dynamics, Ph.D. course, 9 ECTS

18 two-hours meetings (including today)
1 lecture + 1 exercise session per week
(bi-)weekly homeworks + small final presentation
reference books:
0 notes and specific research papers
1 Easley & Kleinberg, Networks, crowds, and markets, 2010
2 Newman, Networks, Oxford U.P., 2010
3 Levine, Peres & Wilmer, Markov chains and mixing, 2008
4 Draief & Massoulie, Epidemics and rumors in complex networks
5 Aldous& Fill, Reversible Markov chains and random walks
today: broad intro, course overview

Similar courses around the world

MIT grad: http://stellar.mit.edu/S/course/6/sp11/6.986/

MIT undergrad:
http://stellar.mit.edu/S/course/6/fa09/6.207J/

Cornell undergrad (by D.Easley, J.Kleinberg, & E.Tardos)::

http://www.infosci.cornell.edu/courses/info2040/2011fa/

Berkeley grad 1 (by D.Aldous):

http://www.stat.berkeley.edu/ aldous/260-FMIE/index.html

Berkeley grad 2 (by E.Mossel):

http://www.stat.berkeley.edu/ mossel/teach/SocialChoiceNetworks10/index.html

North-Easter U. (by A.Barabasi):

http://barabasilab.neu.edu/courses/phys5116/

Michigan U. (by L.Adamic):

http://open.umich.edu/education/si/si508/fall2008/materials

(Complex) networks

(Large-scale) systems of (simple) interacting units

infrastructure networks: transportation, power, gas, and water

distribution, sewer, Internet

informational networks: WWW, citation networks

social networks: friendships, family ties, Facebook etc.

economic networks: supply chains

financial networks: borrowing-lending nets

biological networks: neural networks, gene/protein interactions

ecological networks: food webs, flocks, ...

Studying (complex) networks

network structure + interaction mechanism

emerging behavior

spread of epidemics and information

design of distributed algorithms

opinion formation, social influence, and learning

network robustness,

cascaded failures, systemic risk

...

Mathematical representation of network structure

(un)directed (weighted) graph G = (V, E)

V = set of vertices (or nodes)

n = |V| < +

E V V = set of edges (or links)

Examples 1

Internet: nodes=routers, edges=direct physical links (und.)

traffic networks: nodes=junctions, links=roads (directed)

actors collaboration: nodes=actors, link same movie (und.)

scientific collab.: nodes=researchers, link=coauthors (und.)

Example 2: political blogs before 2004 US elections

from Adamic and Glance, The Political Blogosphere and the 2004
U.S. Election: Divided They Blog, 2005

Example 3: Family ties in 15th century Florence

from Padgett and Ansell, Robust action and the rise of the
Medici, 1400-1434, 1993

Example 4: High school friendships

from Moody, Race, school integration, and friendship segregation

in America, 2002

Example 5: Sexual contacts

from Newman, The structure and function of complex systems, 2003

Example 6: protein network in yeast nucleus

from Maslov and Sneppen Specicity and stability in topology of

protein networks, 2002

Example 7: Freshwater food web

from Martinez, Artifacts or attributes? Effects of resolution on the

Little Rock Lake food web, 1991

Random networks
network structure + interaction mechanism

emerging behavior
structure of large-scale networks is difficult to describe exactly:
huge or non directly accessible data

Random networks
network structure + interaction mechanism

emerging behavior
structure of large-scale networks is difficult to describe exactly:
huge or non directly accessible data
aggregate properties:

connectivity, diameter / average distance

frequency of subgraphs

degree distribution

statistical approach:

ensemble of graphs

typical properties as n = |V| +

Complex networks

Properties widely observed in empirical studies

1. small world diameter log n
2. high clustering many triangles
3. scale free power law degree distribution

Small world

Milgrams experiment (67): randomly selected group of few

hundreds of people from Omaha (NE). A letter given to each
of them to be delivered to a stock broker living in Boston
(MA). Letter can only be handed to a person know directly.
= 35% letters reach destination, median # of steps: 5.5
= 6 degrees of separation

Albert, Jeong, and Barabasi (99): WWW network, n 800M

average distance of webpages 0.35 + 2.06 log n = 18.59

Power laws
pk

log pk

log k

pk C k
show up in quite different contexts:
percentage of words in a book
percentage of cities of a given size
percentage of people having a certain income
Power law = heavy tails: lots of large cities, lots of rich people
Typically explained by rich-gets-richer mechanisms

Power law networks

dv := #{neighbors of v }

pd :=

1
#{v : dv = d}
n

pd Cd
Empirical studies:

Barabasi and Albert (99): WWW has in 2.1, out 2.7

Faloutos (99): Internet 2.16

actor collaborations: 2.3

Redner (98): citation network: in 2.6, pdout C exp(Kd )

Liljeros (01): # sexual partners per year (in Sweden)

male 3.3, female 3.5
2<3
3

=
=

power law scale free

hdi < +

hd 2 i = +

hdi < + hd 2 i < +

Random graphs 1: Erdos-Renyi

G(n, p) = (V, E)
P ({v , w } E) = p

|V| = n

mutually independent

Random graphs 1: Erdos-Renyi

G(n, p) = (V, E)
P ({v , w } E) = p

n = 100
p = 0.15

|V| = n

mutually independent

Random graphs 1: Erdos-Renyi

G(n, p) = (V, E)
P ({v , w } E) = p

n = 100
p = 0.2

|V| = n

mutually independent

Random graphs 1: Erdos-Renyi

G(n, p) = (V, E)
P ({v , w } E) = p

|V| = n

mutually independent

If p = /n, with high probability as n ,

< 1 = size(largest component) log n

phase

size(largest component) n
small
transition
> 1 =
diam(giant component) log n
world

Poisson degree distribution: pd

limited #(triangles)

d
e d!

NO clustering

NO power law

Random graphs 2: preferential attachment

Barabasi-Albert (99)
1. start from a small given graph n0
2. add a vertex and connect it with d older vertices randomly
chosen with conditional probability their current degree
3. repeat step 2 n n0 times

http://ccl.northwestern.edu/netlogo/models/run.cgi?PreferentialAttachm

Random graphs 2: preferential attachment

p Cd 3 power law

d
diam log n small world

(suitably modified) high clustering

Dynamics over networks

network structure + interaction mechanism

emerging behavior

random walks

linear interactions: averaging / voter model

epidemics: SI, SIR, SIS (contact model)

other: majority model, evolutionary dynamics, games

monotone dynamical systems

...

Random walk

1/3 1/3
1/3

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk

1/2

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk

1/5
1/5
1/5

1/5
1/5

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk

1/3
1/3
1/3

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk

1/5
1/5
1/5

1/5
1/5

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk
1/6
1/6
1/6

1/6
1/6
1/6

P stochastic matrix on V, e.g., Pij = 1/di

P(V (t + 1) = i |V (t) = i ) = Pij
dv
network connected unique stationary distr. (v = P
)
u du

Random walk
1/3
1/3
1/6
1/6
1/6

1/3

1/6
1/6
1/6

how fast does ||P(V (t) = ) || go to 0 ?

when will V (t) hit some other w V?
(t) meet for the first time?
when will independent V (t) and V

Distributed averaging

Gossip model:
every node v has a state xv (t) R
nodes get activated at independent Poisson times
when a node v is activated, it choses a neighbor w at random
and updates its value to xv (t) = (1 )xv (t ) + xw (t )
network connected = convergence to consensus

Distributed averaging

opinioni

0.5

10
tempo

0.5

10
tempo

How is the limit consensus value related to the initial states of the
nodes?
How does the network structure affect the speed of convergence?
What is the effect of heterogeneity of the agents behavior?

Consensus vs disagreement
most mathematical models:
connected network

(asymptotic) consensus
0

Since universal ultimate agreement is an

ubiquitous outcome of a very broad class
of mathematical models, we are naturally
led to inquire what on earth one must
assume in order to generate ...
(Abelson 64)
If people tend to become more alike in their opinions, attitudes,
and behavior as they interact, why do not such differences
eventually disappear? (Axelrod 97)

Gossip model with stubborn agents

G = (V, E) connected
opinions: Xv (t) R

V = AS
S = {stubborn agents}
Xs (t) = xs [0, 1] ,

model propaganda: political parties, media sources, advertising, ...

Gossip model with stubborn agents (contd)

G = (V, E) connected
opinions: Xv (t) R

V = AS
A = {regular individuals}
]0, 1] trust
Tka = t

T a := random clock (rate-1 Poisson)

a chooses b a at random

Xa (t ) = x , Xb (t ) = y

Xa (t) = x + (y x)

Typical sample-path behavior

0.5

xs = 1
xs = 0
Xa (t)
0

s, s S : xs 6= xs

250
t

500

P (NO convergence, NO consensus) = 1

Typical sample-path behavior (contd)

0.5

xs = 1
xs = 0
Za (t) :=

1
t

Rt
0

Xa (u)du
0

250
t

P (Za (t) converges) = 1

500

Voter model
= 1 = voter model:

Xa (t) = Xb (t )

t=0