Models of Network

Price Model
MA 653: Network Science
Instructor: Ashok Singh Sairam
[email protected]
Acknowledgement
• Slides have been heavily borrowed from Prof. Konstantinos
Pelechrinis, School of Information Sciences, University of Pittsburgh

MA 653: Network Science 2

Generative network model
• The random network models we have seen until now are mainly used
to study structural properties of networks
• i.e., some parameters are fixed (e.g., number of nodes, number of edges,
degree distribution etc.) and we study the properties of the graph (e.g., path
lengths, component sizes etc.)
• Generative network model: generate synthetic networks that mimic
the structural properties of real-world networks.
• Network models that model the mechanism that drive the network formation
• If the structures resemble real world structures, then this mechanism might
be at work in real networks

MA 653: Network Science 3

Price Model:Origin (1)
• Many real-world networks exhibit power law distributions
• Examples: Internet, WWW and citation network
• What are the underlying processes?
• How did a network come to have such a distribution?
• Price was the first to propose a single and elegant model of network
formation

MA 653: Network Science 4

Price Model:Origin (2)
• Price was studying citation networks
• Nodes are papers. Directed edge if paper 1 cites paper 2. Once a paper is
created it is never destroyed. Paper is added one at a time
• But his model was inspired by the work of Hebert Simon, an economist
• Simon proposed and explanation for the wealth distribution
• People who have more money already, gain more at a rate proportional to
how much they already have
• This can lead to power law distribution for the wealth
• “Rich-get-richer”, “cumulative advantage”, “preferential attachment”

MA 653: Network Science 5

Price model: Idea
• Price adopted with modifications Simon’s model in the context of
(citation) networks
• Every new paper (node of the graph) cites on average c other papers (c is
thus, the average out degree)
• This newly appearing paper cites previously published papers at random
NOT uniformly at random
• At random with probability proportional to the number of citations those
previous papers have
• Simplified model does not consider relevance, originality, difference between
research articles and reviews

MA 653: Network Science 6

Price Model:Adjustment
• The above cannot be precisely true
• A newly published paper has zero citations  a strict proportionality rule
would give 0 citations to this paper from then on

• Price added a constant factor a > 0 to the citation probability

• Some citations come for “free”
• An alternative interpretation of this factor is that a part of the papers that one
cites are chosen uniformly at random, while the rest are chosen with a
probability proportional to the citations these other papers already have

MA 653: Network Science 7

Price model: Initialization
• How to initialize the model?
• At the limit of a large network the initialization does not alter any of the
predictions
• Hence, we can start with a few papers that have zero citations each
• In summary,
• Price’s model is a growing network
• New nodes are added continually and never removed
• Average out-degree is c
• The citations a paper receives are proportional to its in-degree and constant
factor a [Price considered a = 1]
• The model allows for multi-edges but as with random graphs that do not
affect the results at the limit of large network

MA 653: Network Science 8

Price model: properties
• Price model creates acyclic graphs
• This aligns with the initial goal to model citation networks
• What about other directed networks that are not acyclic? E.g., the Web

• What is the in-degree of vertices in the price model?

• We denote for simplicity the in-degree of vertex i as qi
• Let pq(n) be the fraction of vertices that have in-degree q when the total
number of vertices in the network is n
• What happens when a new vertex is added to the network?

MA 653: Network Science 9

Price Model: degree distribution
• This new vertex will create new citations/edges
• The probability that a new edge points to a node i is proportional to
qi+a: qi + a qi + a qi + a where 𝑞 is the average in-degree
= =
å (q + a)
i i n q + na n(c + a)
assuming 𝑞 = c
• This new vertex has c citations  the expected number of new in edges of
node i is c times the above quantity
• The expected number of new edges pointing to vertices of in-degree
q is: q+a c(q + a)
npq (n)´ c ´ = pq (n)
n(c + a) c+a

MA 653: Network Science 10

Price Model: degree distribution
• Let’s calculate what is the number of vertices with degree q after the
addition of this new vertex
• A node with in-degree q can appear because a node with degree q-1 received
a new citation
• However, a node with in-degree can disappear because it received a new
citation and hence, its in-degree is now q+1
• Hence the expected number of nodes with in-degree q when we add
one vertex (i.e., total nodes n+1) is:
c(q -1+ a) c(q + a)
(n +1)pq (n +1) = npq (n) + pq-1 (n) - pq (n)
c+a c+a
#nodes previously of in-degree q #nodes gained #nodes lost
MA 653: Network Science 11
• The master equation we saw previously does not hold true as is for
the case of q=0
• There are no vertices of lower in-degree so the second term cannot appear
• However, when we add a new vertex its in-degree is by default 0  we add
one node of degree 0
ca
(n +1)p0 (n +1) = np0 (n) +1- p0 (n)
c+a
• Let’s examine the asymptotic behavior of the distribution
• Assuming that it indeed converges for large n, we use the shorthand pq=pq(∞)

MA 653: Network Science 12

• Then we get:

• Re-writing the above equations we can get:

1 a / c q  a 1
p0  , pq  pq 1
a 1 a / c q  a 1 a / c
• After some involved calculations pq can be written as:
B(q + a, 2 + a / c)
pq =
B(a,1+ a / c)
• B(x,y) is the Euler’s beta function

MA 653: Network Science 13

Price Model and power law
• It can further be shown that:
Γ 𝑥 Γ(y)
𝐵 𝑥, 𝑦 = ≅ 𝑥 −𝑦 Γ(𝑦)
Γ(𝑥 + 𝑦)
• Where Γ(y) is the gamma function
• Hence for large x, B(x,y) falls off as a power law with exponent y
• Using the above results we can see that for large values of q the degree
distribution follows:
−𝛼 −𝛼 𝑎
𝑝𝑞 ~ 𝑞 + 𝑎 or 𝑝𝑞 ~𝑞 , where 𝛼 = 2 +
𝑐
• While the Price model does not consider many aspects of the citation
process, even with its overly simplistic assumptions can reproduce
structures similar to the ones emerging in real life

MA 653: Network Science 14