MODELS

OF
NETWORK
FORMATION
Introduction

ASM: Models of Networks 138

 • Degree Distributions:

MODELS OF • Regular Networks

• Random Networks
• Scale-Free Networks
NETWORK
FORMATION • Static Models
• Dynamic Models

ASM: Models of Networks 139

 REAL NETWORKS

Let’s see some examples of real networks

SOURCE: “Chains of Affection: The Structure of SOURCE: “The political blogosphere and the 2004 U.S.
Adolescent Romantic and Sexual Networks”, election: divided they blog”, Lada A. Adamic, and N.
Peter S. Bearman, James Moody, and Katherine Glance
Stovel.

ASM: Models of Networks 140

 REAL NETWORKS
• The Jefferson High school links create a network that is
• Almost bipartite: considering male and female as two distinct types of nodes, with few exceptions, the links are only
between the groups and not within
• The degree distribution closely match that of a network that has been created uniformly at randomly (more on this
later)
• There exists a very large component, known as a Giant Component
• The network is tree-like, with not too many cycles and only one very large cycle in the giant component

• The Blogsphere in 2004 shows

• A network with two well defined subgroups corresponding with the political views of the blogs
• The network highly depends on homophily, i.e. the blogs are highly biased to link blogs with the same view. This bias is
known as inbreeding homophily

The existence of so different networks, suggests the need to explore network models that can explain why and how different
networks are created.

ASM: Models of Networks 141

 MODELS OF NETWORKS

• Having summaries as the degree, or the average path length is good to gain knowledge of the network. However,
since networks differ significantly on the number of links, and nodes, quantities as the average degree will vary across
the data sets.

• It is better to have an overall description of the distribution of values of the degrees. This is the degree distribution of
the network

• In the discrete formulation we may think of the degree distribution as returning the fraction of nodes with degree k.

• In general, we take them as a probability distribution describing the probability that if we select a random node in the
network, it will have degree k

• When we talk about random networks, or scale free networks we refer to the degree distribution that describes them
better, i.e., the models of networks correspond to the different processes that give rise to the different probability
distributions.

ASM: Fundamentals of Networks 142

 REGULAR NETWORKS
• These are a degenerate limiting case in which every node is connected to the same number of other nodes, i.e. a circular
network like the one in the figure

The degree distribution for the example on the left, where each
node has degree 2 is

1 if k = 2
𝑃 𝑘 =ቊ
0 otherwise

This network represents an extreme case of order.

ASM: Models of Networks 143

 RANDOM NETWORKS

• These models are in the opposite extreme in the order-disorder axis, as they represent the disorder.

• There are two equivalent models due to Erdös and Renyì, we will only consider the G(n,p) model, in which there are N
nodes and each link between nodes is present with a probability p, giving rise to a binomial distribution

𝑘𝑚𝑎𝑥 𝑘 𝑁−1 𝑘
𝑃 𝑘 = 𝑝 (1 − 𝑝)𝑘𝑚𝑎𝑥 −𝑘 = 𝑝 1−𝑝 𝑁−𝑘−1
𝑘 𝑘

• Note that the number of nodes is fixed, and hence this is a static model

• They are the model used as reference in all the comparisons to detect any structure: degree correlation function,
modularity, assortativity,…

ASM: Models of Networks 144

 RANDOM NETWORKS
Examples of random networks

ERDÖS-RENYÌ MODEL WITH N=30 AND P=0.03 ERDÖS-RENYÌ MODEL WITH N=30 AND P=0.1

ASM: Models of Networks 145

 SCALE FREE NETWORKS
• These models have as degree distribution a power law

𝑃 𝑘 = 𝐴𝑘 −𝛾

• This type of models have fat-tailed distributions, i.e. there are more nodes with low degree and with high degree than
there would be if the links were created at random.

• There may also appear anomalies depending on the value of the degree exponent, 𝛾.

• This degree exponent can be estimated using a linear regression (not too recommended) and with a maximum likelihood
method (better).

• These networks have a creation mechanism that adds nodes at each iteration, hence they are dynamic processes.

ASM: Models of Networks 146

 SCALE FREE NETWORKS
Examples of scale-free networks

BARÁBASI-ALBERT MODEL WITH N=30 BARABASI-ALBERT MODEL WITH N=30

ASM: Models of Networks 147

 STATIC MODELS

Random Networks

ASM: Models of Networks 148

 • Models
• Poisson Networks
• Small-Worlds
• α and β-model

• Properties
• Maximum and Minimum degrees
• Diameter,
• Clustering Coefficient
STATIC MODELS • Average Path Length

• Thresholds and Phase Transitions

• Connectedness

• Giant Components

ASM: Models of Networks 149

 POISSON RANDOM NETWORKS

• A well-known limit of the binomial distribution is the Poisson distribution, which occurs when the sample size is
large, while the probability of success is small, keeping the product 𝑛𝑝 small.

• This limit, known as the rare events limit, will correspond to sparse networks with a large number of nodes.

• Since the average degree of the random network can be written as

𝑘ത = 𝑘𝑚𝑎𝑥 𝑝 = (𝑁 − 1)𝑝
Then the degree distribution will be

𝑘
𝑛−1 𝑝
𝑃 𝑘 = 𝑒− 𝑛−1 𝑝
𝑘!

ASM: Models of Networks 150

 CLUSTERING COEFFICIENT

• The clustering coefficient is the probability that two neighbours of a node are connected. However, in
a random network, the probability that any two nodes are connected is p, the same for any pair of
nodes, then we can immediately write that the average clustering coefficient is

𝑘ത
𝐶=𝑝=
𝑛−1

• The value we obtain from this computation will almost always differ from that obtained from a real
network since real networks have, in general, a larger magnitude since links in the neighbourhoods
of other nodes are more likely to be connected than due to random chance

• For random networks this value will be, typically, very small, in contrast with real networks

ASM: Models of Networks 151

 AVERAGE PATH LENGTH

• In the paper by . Fronczak, P. Fronczak, and J.A. Holyst, “Average path length in random networks”
(2018), it is found, using a mean field approximation, that the average path length of a Erdös-Renyì
random network takes the general expression of

log 𝑛 − 𝛾 1
𝑑ҧ = −
ത
log(𝑘) 2

Where 𝛾 is the Euler-Mascheroni constant ~0.5772

• In general, this magnitude is well described in random networks, providing a value that is close to that
of a real network.

ASM: Models of Networks 152

 SMALL WORLDS

• A small world is a network structure that is both highly locally clustered and has a short path length,
two network characteristics that are normally divergent (as we have seen in the random networks)

• The small world phenomenon is that in a network, the distance between two randomly chosen nodes
is short. Since the number of nodes that can be reached from a random node grows geometrically
with the average degree, when this is large enough and averaging over all the pairs of nodes, we have
that the average path length is

log(𝑁)
𝑑ҧ ≈
ത
log(𝑘)

• In Watts constructions, we can see small worlds as those networks that lie between the order and the
randomness, then these are the type of networks that may appear if we consider an interpolation
between a regular network and an ER-network.

ASM: Models of Networks 153

 α-MODEL

It is a model to capture the nature of the social engagement by using a parameter 𝛼 ∈ [0, ∞), that captures the
balance between the constraints of the social structure and the freedom of individual agency

• Cavemen: highly connected groups without connections outside. This world is made from isolated cliques. The
propensity of two disconnected individuals to be connected is very small, but as soon as they share one friend,
this becomes very high.
• Solarian: no one has propensity to connect to anyone in particular. Fragmented world with isolated individuals
that remains the same until two individuals share most of their connections

Watts interpolates between these two worlds by means of the parameter α

1, 𝑚𝑖𝑗 ≥ 𝑘
𝑚𝑖𝑗 𝛼
𝑃𝑖𝑗 = 1 − 𝑝 + 𝑝, 𝑘 > 𝑚𝑖𝑗 ≥ 0
𝑘
𝑝, 𝑚𝑖𝑗 = 0

ASM: Models of Networks 154

 The behavior of the average path length with α is
interesting

α-MODEL • As we increase α, the disconnected components

will be aggregated and the APL increases
With a little bit of help from my friends…
• The APL reaches a maximum where all the graph
is connected

• From that point there is a process similar to a

percolation in ER-models, in which randomness
is increased for a fixed number of edges

ASM: Models of Networks 155

 β-MODEL The degree distribution is

ത𝑘
𝑘 ത
min(𝑘− , ) ത
𝑘
In the Watts-Strogatz β-model we introduce 22 ത ത 𝑘− −𝑁
the probability of rewiring to nodes ത
𝑘/2 𝑘
𝑁 2−𝑁
(𝛽 𝑘/2) 2
ത
𝑃 𝑘 = ෍ 1−𝛽 𝛽 𝑒 −𝛽𝑘/2
𝑁 𝑘ത
𝑁=0 𝑘− −𝑁 !
2
• Start with a regular network and choose
each node in order

• First, we require with probability β all the

nearest neighbours links

• When the nearest neighbours are done

for all the nodes, go to the second order

• Proceed until all the links have been β increases

considered for rewiring

ASM: Models of Networks 156

 β-MODEL

There is no sociological justification for this

model, then

However, it can be seen that that both, α and β,

are different manifestations of the same
mechanism.

The WS model gives a reason for the average

clustering coefficient, but still a wrong degree
distribution

ASM: Models of Networks 157

 PHASE TRANSITIONS IN RN

• In random networks some different phase transitions occur. These represent moments in their
evolution with respect to any of their parameters in which the network structure changes as the
formation process is modified

• To capture the dynamics of a random network, we can begin with N isolated nodes and go adding the
links according to the probability p. This leads to the study of the evolution of a random network that
will depend almost exclusively on the value of the average degree at each stage

• The evolution we present corresponds to an interpolation between the limiting cases:

𝑝 = 0, where the largest component has size 1 (each isolated node)

𝑝 = 1, where the largest component has size n (all the nodes of the network)

ASM: Models of Networks 158

 PHASE TRANSITIONS

Rembember that in random networks, changing the link probability implies changing the average degree, then

ASM: Models of Networks 159

 GIANT COMPONENTS

• In studying the properties of random networks Erdös and Renyì found that if the probability of a link is larger than

log 𝑁
𝑁

Then the network is connected with probability tending to 1, i.e. the fraction of nodes with no links tends to 0.

• It is a convenction to say that if the number of nodes in a component is fewer than 𝑛2/3 /2 the component is small, and large
otherwise.

• The term Giant Component refers to the only largest component (if there is one). In this component its size grows in
proportion to 𝑛. Formation of a Giant Component is a phase transition where the size of the largest component goes through
a sudden change where the average degree goes from 0 to not-0, growing up to 1.

ASM: Models of Networks 160

 MAXIMUM AND MINIMUM

• In the Poisson limit it is easy to find the

expected degree of the largest node, known
as the upper natural cut-off. Defining 𝑘𝑚𝑎𝑥
such that in a network with n nodes, at most
one node has degree larger than that, then 1
• Maximum: 1 − 𝐹 𝑘𝑚𝑎𝑥 =
𝑛
• We can also the same for the minimum
expected degree 𝑘𝑚𝑖𝑛 , by defining it as the 1
• Minimum: 𝐹 𝑘𝑚𝑖𝑛 − 1 =
value such that we will find, on average, at 𝑛
most one node with degree smaller than it.
ത
For a world like network with 𝑁 = 109 and 𝑘~1,000 we have
𝑘𝑚𝑎𝑥 ~1,200 and 𝑘𝑚𝑖𝑛 ~800

Which are the implications of these values??

ASM: Models of Networks 161

 EVOLUTION OF RN

Subcritical Critical Supercritical Connected

1 1 1 log(𝑁)
Probability < > >
𝑁−1 𝑁−1 𝑁−1 𝑁

Average Degree (0, 1) 1 >1 > log(𝑁)

Giant Component No No Yes Yes

Structure Trees May have Cycles GC with tres and cycles GC with cycles

Size of the Largest 2/3

𝑘ത 1
~ log(𝑁) ~𝑁 ~ − 𝑁 𝑁
Component 𝑁 𝑁

ASM: Models of Networks 162

 RANDOM NETWORKS

Degree Distribution Average Path Length Clustering Coefficient

Erdös-Rényi Model Wrong Right Wrong

Watts-Strogatz Model Wrong Right Right

ASM: Models of Networks 163