1

Small World Networks

Jean Vaucher
Ift6802 - Avril 2005
ift6802 2
Contents
Pertinence of topic
Characterization of networks
Regular, Random or Natural
Properties of networks
Diameter, clustering coefficient
Watts network models (alpha & beta)
Power Law networks
Clustered networks with short paths
Can these short paths be found ?
ift6802 3
Duncan J. Watts
Six degrees - the science of a
connected age, 2003, W.W. Norton.
I read somewhere that everybody on this
planet is separated by only six other
people. Six degrees of separation
between us and everybody on this
planet.
Six degrees of separation by John Guare
ift6802 4
Networks
Networks are everywhere
Internet
Neurons is brains
Social networks
Transportation
Networks have been studied long time
Euler (1736): Bridges of Knigsberg theory of graphs,
which is now a major (and difficult! or almost obvious)
branch in mathematics
ift6802 5
So what is new?
Global interconnections
Internet
Power grids
Mass travel, mass culture

FAILURES
Computer Viruses
Power Blackouts
Epidemics

Modeling & analysis
ift6802 6
Milgrams Experiment
Found short chains of acquaintances linking pairs of
people in USA who didnt know each other;
Source person in Nebraska
Target person in Massachusetts.
Sends message by forwarding to people they knew
personally (who should be closer to target)
Average length of the chains that were completed
was between 5 and 6 steps
Six degrees of separation principle
ift6802 7
Correct question
WHY are there short chains of
acquaintances linking together arbitrary
pairs of strangers???

Or

Why is this surprising
ift6802 8
Random networks
In a random network, if everybody has 100
friends distributed randomly in the world
population, this isnt strange
In 6 hops, you can reach 100
6
people - a
million million > 6,000 million (world pop.)

BUT: our social networks tend to be
clustered.
ift6802 9
Social networks
Not random
But Clustered
Most of our friends come from our
geographical or professional
neighbourhood.
Our friends tend to have the same friends
BUT
In spite of having clustered social
networks, there seem to exist short paths
between any random nodes.
ift6802 10
Social network research

Devise various classes of
networks

Study their properties

ift6802 11
Network parameters
Network type
Regular
Random
Natural

Size: # of nodes
Number of connexions:
average & distribution
Selection of neighbours
ift6802 12
STAR
TREE
GRID
BUS RING
REGULAR Network Topologies
ift6802 13
Connectivity in Random graphs
Nodes connected by links in a purely
random fashion
How large is the largest connected
component? (as a fraction of all
nodes)
Depends on the number of links per
node
(Erds, Rnyi 1959)
ift6802 14
Connecting Nodes
ift6802 15
Random Network (1)
add random
paths
ift6802 16
paths

trees
Random Network (2)
ift6802 17
paths

trees

networks
Random Network (3)
ift6802 18
paths

trees

networks
..

Random Network (3+)
ift6802 19
paths

trees

networks

fully connected
Network Connectivity (4)
ift6802 20
Connectivity of a random graph
1
1
Average number of
links per node
F
r
a
c
t
i
o
n

o
f

a
l
l

n
o
d
e
s

i
n

l
a
r
g
e
s
t

c
o
m
p
o
n
e
n
t

0
D
i
s
c
o
n
n
e
c
t
e
d

p
h
a
s
e

C
o
n
e
c
t
e
d

p
h
a
s
e

ift6802 21
Regular or Ordered Network
ift6802 22
Network measures
Connectivity is not main measure.
Characteristic Path Length (L) :
the average length of the shortest path
connecting each pair of agents (nodes).
Clustering Coefficient (C) is a measure
of local interconnection
if agent i has k
i
immediate neighbors, Ci, is the
fraction of the total possible k
i
*(k
i
-1) / 2
connections that are realized between i's
neighbors. C, is just the average of the Ci's.
Diameter: maximum value of path length
ift6802 23
Regular vs Random Networks
Average number of
connections/node
Diameter
Number of connections
needed to fully connect
few, clustered
Random Regular
fewer, spread
large moderate
many fewer (<2/3)
ift6802 24
Natural networks
Between regular grids and totally
random graphs
Need for parametrized models:
Regular -> natural -> random
Watts
Alpha model ( not intuitive)
Beta rewiring model
ift6802 25
Clustering
Clustering measures the fraction of neighbors of a node that are
connected themselves
Regular Graphs have a high clustering coefficient
but also a high diameter
Random Graphs have a low clustering coefficient
but a low diameter
Both models do match the properties expected from real networks!
Random Graph (k=4)
Short path length
L~log
k
N
Almost no clustering
C~k/n
Regular Graph (k=4)
Long paths
L ~ n/(2k)
Highly clustered
C~3/4
Base metwork is circle
ift6802 26
Small-World Networks
Random rewiring of regular graph (by Watts and Strogatz)
With probability p (or |) rewire each link in a regular graph to a
randomly selected node
Resulting graph has properties, both of regular and random
graphs
High clustering and short path length
FreeNet has been shown to result in small world graphs

ift6802 27
Example: 4096 node ring
Regular graph:
n nodes, k nearest neighbors
path length ~ n/2k
4096/16 = 256
Random graph:
path length ~ log (n)/log(k)
~ 4
Rewired graph (1% of nodes):
path length ~ random graph
clustering ~ regular graph
Small World Graph
K=4
ift6802 28
Small-
world
networks
Beta network
Rewiring probability |
0
1
0
1
L
C
ift6802 29

More exactly . (p = |)
Small world
behaviour
C
L
ift6802 30
Effect of short-cuts
Huge effect of just a few short-cuts.
First 5 rewirings reduces the path
length by half, regardless of size of
network
Further 50% gain requires 50 more
short-cuts
ift6802 31
The strength of weak ties
Granovetter (1973): effective social
coordination does not arise from
densely interlocking strong ties, but
derives from the occasional weak ties
this is because valuable information
comes from these relations (it is
valuable if/because it is not available to
other individuals in your immediate
network)
ift6802 32
Two ways of constructing
ift6802 33
Alpha model
Watts first Model (1999)
Inspired by Asimovs I, Robot
novels
R. Daneel Olivaw
Elijah Baley
Caves of Steel (Earth)
Solaria
ift6802 34
Two extreme types of social
networks
Cavemans world
people live in isolated communities
probability meeting a random person is high if
you have mutual friends and very low if you
dont
Solaria
people live isolated from each other but with
supreme communication capabilities
your social history is irrelevant to your future
ift6802 35
Alpha network
Alpha (o) distance parameter

o=0 : if A and B have a friend in
common, they know each other
(Caveman world)
o= : A & B dont know each other, no
matter how many common friends they
have (Solarian world)
ift6802 36
Number of mutual friends
shared by A and B
L
i
k
e
l
i
h
o
o
d

t
h
a
t

A

m
e
e
t
s

B

Caveman world
Solaria world
o=0
o=
o=1
ift6802 37
Fragmented
networks
Small-
world
net-
works
Alpha network
P
a
t
h

l
e
n
g
t
h

L

o
critical
o
C
l
u
s
t
e
r
i
n
g

c
o
e
f
f
i
c
i
e
n
t

C

L drops because we only count
nodes that are connected
ift6802 38
How about real networks
All nodes in alpha and beta networks are equal in
the sense that the number of connections each
nodes has is not very far from the average
Watts and Strogatz had used normal distribution

Real world is not like that
Sizes of cities, Wealth of individuals in USA, Hubs in
transportation systems
Barabsi and Albert (1999)
Scale-free networks, whose connectivity is defined
by a power-law distribution
ift6802 39
Random Networks
Each node is connected to
a few other nodes.
The number of connections
per node forms a Poisson
distribution, with a small
average of number of
connections per node.
This & three following graphics from:
Linked: The New Science of Networks
by Albert-Laszlo Barabasi; 2002
ift6802 40
Scale-Free Networks
Each node is connected to
at least one other; most are
connected to only one, while
a few are connected to many.
The number of connections
per node forms a hyperbolic
distribution, with no meaningful
average number of connections
per node.
ift6802 41
Random Scale-Free
Scale-free networks are associated with
networks that grow by natural processes
in which the number of nodes increases
with time not just the number of connections.
ift6802 42
Power law phenomena
Average & median are far apart
Whales and minnows
Average from a few large nodes
Median governed by majority of small
nodes



ift6802 43
Performance
Real power law networks also have
short distances
Existence of central backbone of
highly connected HUBS nodes
Similar phenomena noted in
linguistics and economics
Zipf
Pareto
ift6802 44
Zipf's law - linguistics
Zipf, a Harvard linguistics professor,
sought to determine the frequency of use
of the 3rd or 8th or 100th most common
words in English text.
Zipf's law states that the frequency y is
inversely proportional to it's rank r:
Y ~ r
-b
, with b close to unity.

Zipf Presentations
ift6802 45
The Pareto Income Distribution
The Pareto distribution gives the
probability that a person's income is
greater than or equal to x and is
expressed as
| | ( )
parameter shape is
income minimum is
, 0 , 0 , /
k
m
m x k m x m x X P
k
> > > = >
ift6802 46
Vilfredo Pareto, 1848-1923
Italian economist
Born in Paris
Polytechnic Institute in Turin in 1869,
Worked for the railroads.
Pareto did not study economics seriously
until he was 42.
In 1893 he succeeded his mentor, Walras,
as chair of economics at the University of
Lausanne.
QuickTime and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
ift6802 47
Paretos contributions
Pareto optimality.
A Pareto-optimal allocation of resources
is achieved when it is not possible to
make anyone better off without making
someone else worse off.
Pareto's law of income distribution.
In 1906, Italian economist Vilfredo Pareto
created a mathematical formula to describe the
unequal distribution of wealth in his country,
observing that 20% of the people owned 80%
of the wealth.
ift6802 48
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
10000 60000 110000 160000 210000
x
p
(
X
>
=
x
)
Pareto distribution,
m=10000, k=1
0,01
0,1
1
10000 100000 1000000
x
p
(
X
>
=
x
)
log-log plot
Pareto distribution is
said to be scale-free because
it lacks a characteristic length
scale
ift6802 49
Building Power-law networks
It is easy to create PL networks

Build network node by node
Connect new node to an existing
node
Probability of connection proportional
to its number of links
The rich get richer
The poor get poorer
ift6802 50
Structure and dynamics
The case of centrality
centers are in networks
by design (central control, dictatorship)
by non-design (unnoticed critical resources,
informal groups)
or they emerge as a consequence of
certain events
he was at the right place at a right time
clapping in unison
ift6802 51
Further applications
Search in networks
Short paths are not enough
Epidemics: medical & software
Danger of short-cuts
Paths + infectiousness
Infection by ideas
Fads & Economic Bubbles
Individual rationality
Peer pressure

ift6802 52
Getting practical: search in networks
A node may be linked to another
node via a short path but what does it
matter if you cannot find the path?
In alpha and beta networks there is
no notion of distance, therefore
directed searches cannot recognize
shortcuts
Kleinbergs (gamma) networks
(2000)
ift6802 53
Kleinbergs Small-World Model
Embed the graph into an r-dimensional grid (2D in examples)
constant number p of short range links (neighborhood)
q long range links: choose long-range links such that the probability to have
a long range contact is proportional to 1/d
r
Importance of r !
Decentralized (greedy) routing performs best iff. r = dimension of space
(here=2)

r = 2
ift6802 54
Influence of r (1)
Each peer u has link to the peer v with probability proportional to
where d(u,v) is the distance between u and v.

Optimal value: r = dim = dimension of the space
If r < dim we tend to choose more far away neighbors (decentralized
algorithm can quickly approach the neighborhood of target, but then slows
down till finally reaches target itself).
If r > dim we tend to choose more close neighbors (algorithm finds quickly
target in its neighborhood, but reaches it slowly if it is far away).
When r = 0 long range contacts are chosen uniformly. Random graph
theory proves that there exist short paths between every pair of vertices,
BUT there is no decentralized algorithm capable finding these paths
r
v u d ) , (
1
ift6802 55
r
(log scale)
p(r)
(log scale)
increasing
=0

T
y
p
i
c
a
l

l
e
n
g
t
h

o
f

d
i
r
e
c
t
e
d

s
e
a
r
c
h

2
short
paths
cannot
be found
no short
paths
ift6802 56
Influence of r (or )
Given node u if we can partition the remaining peers into sets A
1
,
A
2
, A
3
, , A
logN
, where A
i
, consists of all nodes whose distance
from u is between 2
i
and 2
i+1,
i=0..logN-1.
Then given r = dim each long range contact of u is nearly equally
likely to belong to any of the sets A
i

A
4
A
3
A
2
A
1
ift6802 57
The New Yorker View
When gamma is at
its critical value two,
the resulting
network has the
peculiar property
that nodes possess
the same number of
ties at all length
scales (in 2D world)
ift6802 58
DHTs (distributed hash tables)
and Kleinberg model

P-Grids
model

Kleinbergs
model

Balanced n-ary search
ift6802 59
More hierarchy
Kleinbergs model has only one distance
measure, geographical (2D)
In human society the social distance is
multidimensional
if A is close to B and C is close to B but
in different dimension then A and C can be
very far from each other
violation of the triangle inequality
but multidimensionality may enable messages
to be transmitted in networks very efficiently
ift6802 60
Watts et al (2002) search in social
networks
Searchable
networks
H
1
10
0
6
o
Kleinberg
condition
o= homophily, the
tendency of like to
associate with like

H=number of dimensions
along which individuals
measure similarity
ift6802 61
Small Worlds
& Epidemic diseases
Nodes are living entities
Link is contact
3 States
Uninfected
Infected
Recovered (or dead)

ift6802 62
Epidemic diseases
Level of infectiousness needed to start an epidemic varies
with presence of shortcuts
In regular grid, disease may die out due to lack of victims
In small world, pandemics are facilitated
SRAS
Mad cow disease in England

0
Fraction of random shortcuts
1
Threshold
infectiousness
ift6802 63
Failures in networks
Fault propagation or viruses
Scale-free networks are far more resistant
to random failures than ordinary random
networks
because of most nodes are leaves
But failure of hubs can be catastrophic
vulnerable or targets of deliberate attacks
which may make scale-free networks more
vulnerable to deliberate attacks
Cascades of failures
64
Back to Social Networks
ift6802 65
Spread of ideas
Messages in social networks
Fads & fashions
Body piercing, baseball caps
Harry Potter, Amlie Poulin
Innovation, scientific revolutions
Solar-centric universe
Plate tectonics
Is it like the spread of disease ?

ift6802 66
Effect of peers & pundits
Peoples decisions are affected by
what others do and think
Presure to conform ?

Efficient strategy when insufficient
knowledge or expertise
Ex: picking a restaurant

ift6802 67
Economic models
Selfish agents
Individual rationality
Markets
Equilibrium ???
Many agents are trend followers
Speculation crashes

ift6802 68
Social Experiments
Factors which affect decisions
Milgram
Asch
ift6802 69
Stanley Milgram (1933-1984)
Controversial social psychologist
Yale & Harvard
Small world experiment, 1967
6 degrees of separation
Obedience to authority - 1963
ift6802 70
Validity of Milgrams experiment
Global connectivity ?
US: Omaha Boston stockbroker
Only 96 valid subjects (out of 300)
100 from Boston
100 big investors
96 picked at random in Nebraska
Success? 18 out of 96
Other experiments:
3 out of 60
Worse.

ift6802 71
Conformity
Other presentation
ift6802 72
Threshold models of decisions
Number of infected
neighbors
1
P
r
o
b
a
b
i
l
i
t
y

o
f

i
n
f
e
c
t
i
o
n

0
Fraction of neighbors
choosing A over B
1
P
r
o
b
a
b
i
l
i
t
y

o
f

c
h
o
o
s
i
n
g


o
p
t
i
o
n

A

0
Critical
Threshold
Standard disease spreading
model
Social decision making
ift6802 73
Global Cascades
Idea catches on.
ift6802 74
Fin