Team Control Number

For office use only For office use only

T1 ________________
25318 F1 ________________
T2 ________________ F2 ________________
Problem Chosen
T3 ________________ F3 ________________
T4 ________________ C F4 ________________

2014
Mathematical Contest in Modeling (MCM/ICM) Summary Sheet

Who are the 20%?

The famous ‘80-20’ rule states that the 80% influence is caused by the 20% for many
events. This principle also applies to the network science: only few nodes have a
significant influence and impact to the whole network. In our paper, A Relation
Distance Model and an Authority-Popularity Evaluation Model are employed to
measure of the 20% and analysis of its influence.

For requirement 1 & 2, we construct the undirected co-author network based on the
511 order relationship matrix. A Relation Distance Model is proposed on the basis of
SNA technique. It combines three centrality indexes as a measure vector to calculate
the ‘distance’ with the most influential node. Another measure (Eigenvector
centrality) which takes both degree and the influence of co-authors into
consideration outputs a new rank. Validation of the model is discussed by
comparing the two ranking of the top 15 authors in Erdos1 network, we find ALON,
NOGA M. is the most influential person in the Erdos1 network. The degree
distribution of the Erdos1 network is proved to approximately be power-law
distribution, which indicates it is a scale-free network.

For requirement 3, an Authority-Popularity Evaluation Model is established to

analyze the depth and the width of the influence of nodes. An Authority Index is
calculated by our Modified PageRank Algorithm (including two steps: i nitialization
and iteration) to measure the depth of impact. A Popularity Index is defined as
citation per year to reflect the width of influence. A set of 24 papers citation directed
network with weight added to nodes is constructed to implement the algorithm. The
final ranking of the papers is obtained by combining Authority-Popularity Indexes.
For requirement 4, a set of 15 actors’ co-star bidirectional network with co-star movie
as links is constructed. The iteration process is refined as the weight added to links
instead of nodes.

For requirement 5, we discuss two characteristics of scale-free network: Growth and

Preferential attachment. The philosophy of dynamics of scale-free network is
revealed as the 'Matthew effect'. The prior method to boost influence is proposed:
finding the shortest links to the most influential author. Cooperating with the 80%
(more approachable relatively) with low closeness centrality step by step and finally
co-author with the key figure in the field.

A sensitivity analysis is conducted to study the robustness of our algorithm to Damp

Coefficient and the result shows a good stability. Strength and weakness of our
models is also discussed.

更多数学建模资料请关注微店店铺“数学建模学习交流”
https://k.weidian.com/RHO6PSpA
Team # 25318 Page 1 of 20

Who are the 20% ?

1. Introduction

In this paper, we will introduce two proper methods to analyze influence and
impact in research networks and other areas of society.

The Relation Distance Model is proposed based on SNA technique. It is a

combination of three centrality indexes. Another measure (Eigenvector central-
ity) which takes both degree and the influence of co-authors into consideration
outputs a new rank. Validation of the model is discussed by comparing the two
ranking of the top 15 authors in Erdos1 network.

The Authority-Popularity Evaluation Model is established in section 5 to an-

alyze the depth and the width of the influence of important nodes. The Authori-
ty Index reflects the depth of impact and the Popularity Index reflects the width
of influence. A set of 24 papers citation network with weight is constructed to
implement the algorithm and the results are discussed. In order to analyze a co-
star network for Requirement 4 , a refinement to the model based on Modified
PageRank Algorithm is proposed in subsection 5.3.

The philosophy of the scale-free network is revealed and the shortcut to boost
influence based on our model is proposed in section 6.

2. Basic Assumptions

In this section, we discuss several assumptions we have made and rationale

for making these assumptions.

Assumption 1. We assume that the strength of co-authorship between two

arbitrary Erdos1 authors is the same.

We believe that the accurate strength level of co-authorship between two ar-
bitrary Erdos1 authors is hard to be measured by some criterion. For the sake
of simplification, in our case, if two authors co-author a paper , then the co-
authorship index is ’1’, if not , it is ’0’.

Assumption 2. We assume that the significance of a research paper is deter-

mined by both its citations and publishing date , also, the influence of the journal
should be considered.
Team # 25318 Page 2 of 20

We believe that the more often a paper is cited or the earlier a paper is pub-
lished, the more influential it is. And also, a high influential journal contributes
more to the influence of a paper.

Assumption 3. We assume that the quality of a movie is determined by its

IMDb rating[1] . The popularity of a movie star is measured by the number of
Google search results.

3. Definitions

Definitions of symbols employed in this paper are listed in Table 1.

Table 1: Symbol table.

Variable Description
Relation Distance Model
x The index of member nodes
aix Relation strength
dix Relation distance
Cd (x) The degree centrality of node x
Cb (x) The betweenness centrality of nodex
Cc (x) The closeness centrality of node x
gij (x) The shortest path between two other nodes i and j
passes through the nodex
lix The length of the shortest path connecting node i and x
n The total number of nodes in a network
−→
Ax The vector contains the three centrality measures
−→
AC The ideal vector of the node which has the most
significant influence within the network
DC (x) The Euclid distance defined to measure the influence
and impact of node x
Eigenvector centrality
Ex The eigenvector centrality value of node x
B = (bij ) Adjacent matrix of a network
c Proportional constant
Revised PageRank Algorithm
P Rx (0) Initial PageRank value of node x
s Scale constant for the Revised Algorithm
Gij The relationship matrix
D The coefficient matrix
Team # 25318 Page 3 of 20

4. Models for Requirement 1 and 2

4.1 Data Preprocessing

Before presenting our models, we would like to address the preprocessing

work we have done to the data.

Step 1. We extract the 511 Erdos1 authors from over 18,000 lines of raw data
in Erdos1 file , which is a easy task by eliminating names without a date followed
by in Microsoft Excel 2010.

Step 2. In order to obtain the relationship among the 511 Erdos1, we leave
out the Erdos2 from the list of co-authors of the 511 Erdos1 by using library
function ’countif()’ in Microsoft Excel 2010.

Step 3. We construct a 511 order co-authorship matrix by executing a pro-

gram in MATLAB@2012b. For the sake of description, we give each Erdos1 au-
thor an ID number by the rule:

’1’ stands for ABBOTT, HARVEY LESLIE

’2’ stands for ACZEL, JANOS D.

...

’511’ stands for ZIV, ABRAHAM

4.2 Relation Distance Model based on SNA

4.2.1 Overview

In order to build the Erdos1 network and analyze its properties, the social
network analysis (SNA) technique is employed. Social Network Analysis refers
to methods used to analyze social networks structures made up of individu-
als called ’nodes’, which are tied (connected) by one or more specific types of
interdependency. In our case, the Erdos1 authors are viewed as nodes and co-
authorship (obtained from subsection 4.1) as links among them.
Team # 25318 Page 4 of 20

Figure 1: The co-author network of the Erdos1. For the sake of observability, we choose the
top 50 Erdos1 authors in the priority list in subsection 4.2.2.

4.2.2 Methodology

In the following, we will finish the steps to build up and validate the model.
−
→
Step 1. Calculate the centrality measure vector Ax of the nodes in the net-
work.

According to Freeman’s research[1979], there are three popular centrality

measures- degree centrality Cd (x), betweenness centrality Cb (x) and closeness
centrality Cc (x) . What’s more, they can be used to identify ’masters’ who have
significant influence or impact in a network. These values are defined as below:

Degree centrality is defined as:

n
P
Cd (x) = aix
i=1

where n is the total number of nodes in a network, and aix is a variable indicat-
ing the weighted number of co-authorship between nodes x and i. According to
Assumption 1, in our Erdos1 network, aix = 1 or 0 , for all i.

Betweenness centrality is defined as:

Team # 25318 Page 5 of 20

n P
P n
Cb (x) = gij (x)
x j

where gij (x) indicates whether the shortest path between two other nodes i and
j passes through the node x.

Closeness centrality is defined as:

n
P
Cc (x) = lix
i=1

where lix is the length of the shortest path connecting nodes i and x. The shortest
paths are calculated based on the Floyd algorithm.

The centralities above describe different characters of nodes in a network.

• Degree centrality shows the number of nodes’ connections, which also re-
flects connectivity of nodes in a network. Nodes with high connectivity
can be viewed as more influential.

• Betweenness centrality shows the number of the shortest paths that pass-
ing by certain node. It also reveals the dependency of a node from other
nodes. Obviously, if a node is dependent by other nodes quite much, the
node must be very important for the whole network, That is, large be-
tweenness centrality value of a node is equivalent to its high importance
in the network.

• Closeness centrality actually measures how far away one node is from oth-
er nodes. Apparently that small closeness value of a node reflects its high
importance.

According to the above, we define a vector which contains the three measures
as the form below:

−
→ Cd (x) Cb (x) Cc (x)
Ax = ( , , )
maxCd (x) maxCb (x) maxCc (x)
−
→
Ax is called ’measure vector’. It can also be represented in another from below:
−
→
Ax = (Ax1 , Ax2 , Ax3 )

−
→
where Axi (i=1,2,3) stands for element i in the vector Ax . The three elements are
all divided by their maximum values to be normalized, separately. According
to the definition of three centralities , it is obvious that Ax will get its optimal
value when degree(Ax1 ) and betweenness (Ax2 ) get to their largest value 1 and
closeness (Ax3 ) gets to its smallest value 0.
Team # 25318 Page 6 of 20

Therefore, the ideal model of an author who has the most significant influ-
−
→
ence within the network will have his/her own measure vector Ax as (1, 1, 0).
−→
The ideal vector will be defined as AC :
−→
AC = (AC C C
1 , A2 , A3 ) = (1, 1, 0)

Step 2. Calculate the ’distance’ among the member nodes.

−
→
The measure vector Ax can be used to calculate the ’distance’ among the
−→
member nodes. We define the Euclid distance of Ax from the node x to the ideal
−→
vector AC as DC (x). Based on the idea of ranking method from TOPSIS algo-
rithm [ S. Mahmoodzadeh 2007 ], we call the ’distance’ as Influence and Impact
Distance, which is defined as below:
q
DC (x) = (Ax1 − AC 2 C 2 C 2
1 ) + (Ax2 − A2 ) + (Ax3 − A3 )

Step 3. Generate the influence priority list.

This distance is the key to determine who in the Erdos1 network has signif-
icant influence within the network. A further distance apparently indicates a
lower possibility of being an author who has significant influence. The Erdos1
authors will be arranged into an Influence and Impact priority list. The order of
the list is arranged according to the value of DC (x). Node with smaller DC (x)
value will rank higher in the priority list since DC (x) is the distance from the
node to an ideal ’most significant’ node.

4.2.3 Results and Analysis

The three-dimensional graph is shown in Figure 2 , in which 511 points stand

for 511 Erdos1 authors’ measure vectors correspondingly .

The top 15 authors of the Influence and Impact priority list are shown below
in Table 2. For the sake of observability, we choose the top 50 Erdos1 authors in
the impact priority list to draw the network in UCINET, as is shown in Figure 1.

We can see from Table 2 that HARARY, FRANK* has the most significant
influence within the co-author network. When it comes to a specific university,
department, or a journal, the relationship matrix data should be collected for the
network. The centrality index can be calculated and a ranking can be obtained
from our model. Weight of nodes or links may be added if necessary, and we
will discuss these network in section 5.
Team # 25318 Page 7 of 20

Figure 2: Measure vectors of 511 Erdos1 authors, the big red point(1,1,0) stands for the ideal
vector of the most crucial one.

Table 2: Top 15 most influential authors within the Erdos1 network.

Ranking DC (x) ID Name of authors

1 0.1798 187 HARARY, FRANK*

2 0.2927 438 SOS,VERA TURAN
3 0.2962 10 ALON,NOGA M.
4 0.3400 165 GRAHAM,RONALD LEWIS
5 0.3557 148 FUREDI,ZOLTAN
6 0.3587 44 BOLLOBAS,BELA
7 0.4793 479 TUZA,ZSOLT
8 0.5022 355 POMERANCE,CARL BERNARD
9 0.5259 449 STRAUS,ERNST GABOR*
10 0.5394 341 PACH,JANOS
11 0.5778 180 HAJNAL,ANDRAS
12 0.6531 378 RODL,VOJTECH
13 0.6998 440 SPENCER,JOEL HAROLD
14 0.7440 249 KLEITMAN,DANIEL J.
15 0.7653 399 SARKOZY,ANDRAS

Eigenvector centrality
For the second question in Requirement 2 ,now we consider who has pub-
lished important works or connects important researchers within Erdos1.

The importance of a node is both determined by the number of its neighbor

nodes (its degree) and the importance of its neighbor nodes.In graph theory and
network analysis, centrality of a vertex measures its relative importance within
a graph[4] .
Team # 25318 Page 8 of 20

The definition of eigenvector centrality is :

n
X
Ei = c bij Ej
j=1

where c is the proportional constant, B = (bij ) is the adjacent matrix of the

network.

Calculate the 511 nodes’ eigenvector centrality value in UCINET , we get the
top 15 nodes as is shown below in Table 3.

Table 3: Top 15 connecting important researchers within Erdos1.

Ranking E value ID Name of authors

1 0.26 10 ALON, NOGA M.

2 0.234 378 RODL, VOJTECH
3 0.209 44 BOLLOBAS, BELA
4 0.204 165 GRAHAM,RONALD LEWIS
5 0.201 148 FUREDI,ZOLTAN
6 0.187 479 TUZA,ZSOLT
7 0.179 440 SPENCER, JOEL HAROLD
8 0.176 177 GYARFAS, ANDRAS
9 0.173 462 SZEMEREDI, ENDRE
10 0.161 128 FAUDREE, RALPH JASPER, JR.
11 0.158 287 LOVASZ, LASZLO
12 0.151 78 CHUNG,FAN RONG KING(GRAHAM)
13 0.151 341 PACH, JANOS
14 0.146 261 KOSTOCHKA, ALEXANDR V.
15 0.146 326 NESETRIL, JAROSLAV

We can see from Table 3 that ALON, NOGA M. is the person who connects
more important researchers within Erdos1 than others .

4.3 Validation of the Model

Comparing the two kinds of ranking methods, the majority of the authors in
the first table are also in the second table. However, some ’specific’ authors in
the first table cannot find their name in the second top list. Why? Next, we will
focus our discussion on the problem by the case of HARARY, FRANK*.

The degree of HARARY, FRANK* is the maximum 44, which means he coau-
thored with 44 other Erdos1 authors. When we study the relatively most influ-
ential authors in the network, we discover that only 30 authors are the ’master’
within the network. That is to say, most of the co-authors of HARARY, FRANK*
Team # 25318 Page 9 of 20

are not high influential authors. So, when we consider both degree and the in-
fluence of his/her co-authors, authors whose degree ranks high may not be the
’master’ within the network. So, the weakness of the Relation Distance Model is
being susceptible to large degree value.

4.4 Properties of the Erdos1 network

In subsection 4.2, we have got each node’s degree centrality. So, the degree
distribution of the network can be obtained. Now, we focus our discussion on
the degree distribution of the Erdos1 network. According to an algorithm com-
bined maximum-likelihood fitting methods with goodness-of-fit tests proposed
by Clauset [2007] , we discover that the distribution of degree centrality k in the
Erdos1 network is approximate a power-law distribution, the power-law expo-
nent is estimated as 1.6.

P (K > k) ∼ k −1.6

The degree distribution of the Erdos1 network and the approximative power-
law distribution is shown below in Figure 3.

Figure 3: The degree distribution of the Erdos1 network and the approximative power-law
distribution.

According to the idea of Barabsi [1999], we consider the Erdos1 co-authoring

network as a scale-free network. In the scale-free network, most nodes have
small degree value , only very few nodes have large degree value, the degree
Team # 25318 Page 10 of 20

distribution of the scale-free network appears to be approximately power-law

distribution.

Some other properties are shown below:

• Overall graph clustering coefficient of the network is 0.343

• Average distance (among reachable pairs) of the network is 3.825

5. Authority-Popularity Evaluation Model

5.1 The Citation Network for Requirement 3

In Requirement 3,we construct a directed network connection graph with

weight. The 16 foundational papers listed in the attached list (NetSciFounda-
tion.pdf ) and 8 additional paper we discover are considered as the nodes in the
Citation network, and the citation relationship as the links. The Citation net-
work is shown as Figure 4. The additional papers we discover are listed as
below:
No.17: Newman, M. E. J., Strogatz, S. H., and Watts, D. J. Random graphs with arbitrary
degree distributions and their applications, Phys. Rev. E 64, 026118 ,2001.

No.18: Bollob’as, B., Random Graphs, Academic Press, New York, 2nd ed.,2001.

No.19: Holland, P. W. and S. Leinhardt . An exponential family of probability distributions

for directed graphs. Journal of the American Statistical Association 76, 33-65 ,1981.

No.20: Snijders, T. A. B. Markov chain Monte Carlo estimation of exponential random graph
models. Journal of Social Structure 3(2), 2002.

No.21: Watts, D. J., Small Worlds , Princeton University Press, Princeton ,1999.

No.22: M.Barthlmy, The architecture of complex weighted networks, Proceedings of the

National Academy of Sciences, 2004.

No.23: Luis A Nunes Amaral, Antonio Scala, Marc Barthlmy, H Eugene Stanley. Classes of
small-world networks, Proceedings of the National Academy of Sciences,2000.

No.24: M. Barthlmy, Small-world networks: Evidence for a crossover picture Physical Re-
view Letters, 1999.

• Explanation 1: Let paper A is cited by paper B, then the direction of the

link between them is from B to A (for the implementing of PageRank Al-
gorithm).
Team # 25318 Page 11 of 20

Figure 4: The directed citation network with weight added to nodes.

• Explanation 2: According to Assumption 2, we define the weight of a node

as the impact factor of the journal [7] in which the paper published.

• Explanation 3: For the sake of description, we give each paper an ID. ’1’
for the first paper in NetSciFoundation.pdf, ’2’ for the second ,’17’ for the
first paper in the additional list, etc.

5.2 Authority-Popularity Evaluation Model based on the revised

PageRank Algorithm

The aim of Requirement 3 is to determine the relative influence of the papers.

In this task, we develop an Authority-Popularity Evaluation Model which is
a combination of the revised PageRank Algorithm and Normalized Influence
Factor. In our model, the PageRank value and the Influence Factor reflects the
depth and the width of the impact, respectively.

The Modified PageRank Algorithm

The basic idea of the PageRank Algorithm is that the importance of a node is
determined by the quantities and the quality of other nodes pointing to it. The
original PageRank Algorithm is effective for common directed network in ana-
lyzing the importance and impact of the nodes within the network. But when
it comes to Dangling node (node whose out-degree is 0), the Random surfing
will fail as it will be ’trapped’ in the Dangling node forever. We discover that
there is 3 Dangling nodes the Citation Network in subsection 5.2, so a Modified
PageRank Algorithm is proposed to measure the authority of papers in the net-
Team # 25318 Page 12 of 20

work. The Modified PageRank Algorithm includes two steps: Initialization and
Iteration.

Step 1: Initialization

An initial PageRank value (PR value) P Rx (0), x = 1, 2, . . . , n is given to all

nodes in the network, satisfied with
n
X
P Rx (0) = 1
x=1

In our citation network, the initial PR value of each paper is defined as the nor-
malized impact factor of the journal in which the paper is published. The abso-
lute impact factors value (collected from the web [6] ) with ID of each paper are
listed in Table 4. For instance, the weight of paper No.4 ’Emergence of scaling in
random networks’ published in Science is equal to the Impact Factor of Science
31.027.

Table 4: ID of papers and their Weight

No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8

0.322 44.982 0.539 31.027 0.424 3.381 4.375 38.597
No.9 No.10 No.11 No.12 No.13 No.14 No.15 No.16
2.313 9.737 5.952 3.542 31.027 38.597 5.019 3.381
No.17 No.18 No.19 No.20 No.21 No.22 No.23 No.24
2.313 39.265 1.834 2.542 40.664 9.737 9.737 7.943

A 24 order relationship matrix G is constructed by the rule if paper j is cited

by paper i, Gij = 1. Otherwise , Gij = 0. Another matrix is coefficient matrix D,
which is also a 24 order matrix.
   P1 
1 (1)
 1  P1 
  (2) 
D= . . . . . .
 
  
1
1
  P

 (23) 
1 P1
(24)

P 24
P
where (i) = Gji
j=1

Step 2: Iteration

After k times iteration, we get the the PageRank value:

Team # 25318 Page 13 of 20

Figure 5: The flow chart of iteration process

X P Rj (k) 1−q
P Ri (k) = q + i = 1, 2, . . . , n
j
L(j) n

where j represent all nodes that point to node i. q is a constant, defaulted to be

0.85. We implement the iteration based on the equation above until the change
of PageRank is small enough in one step. We set this threshold as 0.0001. We exe-
cute a program for this process in MATLAB@2012b and get the stable PageRank
value P R(k) as is shown in Table 5.

Table 5: The rank of papers by the stable PageRank value.

Ranking P Ri (k) ID Ranking P Ri (k) ID

1 0.01575 No.18 13 0.00094 No.19

2 0.01217 No.14 14 0.00060 No.11
3 0.00937 No.4 15 0.00060 No.23
4 0.00784 No.2 16 0.00051 No.24
5 0.00493 No.1 17 0.00051 No.20
6 0.00278 No.17 18 0.00051 No.13
7 0.00183 No.21 19 0.00047 No.22
8 0.00150 No.3 20 0.00047 No.16
9 0.00142 No.10 21 0.00047 No.15
10 0.00142 No.9 22 0.00047 No.12
11 0.00142 No.8 23 0.00047 No.7
12 0.00111 No.6 24 0.00047 No.5

The result shows that the No.18 paper: Random Graphs, wrote by Bollobs,
has the highest PageRank. So it is the most authoritative paper among 24 paper-
s. Authority reflects the depth of influence. To find the most influential paper
we have to get the width of the influence: popularity. We consider to use cited
quantities to inflect the popularity in the citation network. Since the cited quan-
tities related to the years that paper were published, so we set cited quantities
per year as the parameter, which is shown in Table 6 .

To judge which is the most influential paper, we draw a picture of 24 nodes

(as is shown in Figure 6). For the sake of observability, we set log10 (106 ∗ P Ri (k))
Team # 25318 Page 14 of 20

Table 6: Cited quantities per year of each paper.

ID cited quantities/year ID cited quantities/year

No.14 1355.5 No.8 89

No.4 1256.2 No.1 82.4
No.2 1104.2 No.3 72.4
No.11 965.1 No.13 69.6
No.18 458.2 No.6 54.5
No.21 281.8 No.20 34.6
No.10 211.4 No.16 27.7
No.17 178.2 No.24 19.5
No.23 170.6 No.19 16.9
No.22 142.3 No.15 11.0
No.9 117.2 No.7 0.5
No.12 97.1 No.5 3.125

as ordinate and log10 (10 ∗ citedquantitiesperyear) as abscissa.

Figure 6: Authority-Popularity Diagram for the citation network

In our Authority-Popularity Evaluation Model, we divided Figure 6 into four

regions: (1) high authority, more popular; (2) high authority, less popular; (3)
low authority, more popular and (4) low authority, less popular. It is clear that
the node on the upper right corner is the most influential paper as it is both
the most authoritative and the most popular one. The most influential paper
within the Citation Network in subsection 5.1 is No.14: Collective dynamics of
‘small-world’ networks, wrote by Watts, D. and Strogatz, S.
Team # 25318 Page 15 of 20

5.3 The Co-star Network for Requirement 4

Similar to Erdos number in mathematics, a Bacon number[7] in the film in-

dustry is popular in the late 1990’s. The ’Game of Kevin Bacon’ measures the
shortest path to connect an arbitrary actor/actress with Bacon. Inspired by the
game, we collect a set of 15 famous Hollywood movie stars as nodes and the
co-star movies as links to construct a co-star network(as is shown in Figure 7).
This is also a directed network connection graph with weight, but the difference
is that the weight value is added to the link instead of the node.

• Explanation 1: Let A is the leading actor who is supported by actor B in a

movie, then the direction of the link between them is from B to A.

• Explanation 2: According to Assumption 3, we define the weight of the

link between two stars as the IMDb rating[1] of the movie they co-starred.

Figure 7: The directed citation network with weight added to links.

Since the weight value is added to the link instead of node in co-star net-
work, the PageRank algorithm here has to be improved. By the origin PageR-
ank algorithm, the PageRank value of node i will be distributed to nodes which
are pointed to from i evenly in every step of the iteration. In our improved
PageRank algorithm, how much PageRank value that node i distribute to nodes
that i point to is determined by the weight (IMDb rating[1] of the movie) of the
link. The weight is proportional to PageRank value that is distributed. This is
reasonable because excellent movie can make movie stars more influential. For
example, the weight of the link(i to k) is 3, the weight of the link(i to n) is 4.
3
So the PageRank value that i distribute to k is 3+4 × 1 and the PageRank that i
4
distribute to n is 3+4 × 1, as is shown in Figure 8.
Team # 25318 Page 16 of 20

Figure 8: The diagram for modified algorithm.

So iteration for this algorithm is different. The relationship matrix G in it-

eration for the citation network should be revised to G’ in the co-star network.
We set column j of G as Gj , G0j is defined similarly. It is clearly that if Gij 6= 0,
then G0ij represents a movie. The elements in G0j should be proportional to IMDb
rating of the movie in G0j . For example, We assumed that the movie G2j ’s IMDb
rating is 3, G4j ’s IMDb rating is 4. And

Gj = [0 1 0 1 0]T

So we get
3 6 4 8
G02j = (1 + 1) = G04j = (1 + 1) =
3+4 7 3+4 7

6 8 T
G0j = [0 0 0]
7 7

Other steps are same as in subsection 5.2. We also implement the improved
PageRank algorithm in MATLAB@2012b and result is as below in Table 7.

As is in subsection 5.2, we draw a picture for 15 nodes in the co-star network.

In Figure 9 , we set log10 (the number of Google search results) as abscissa and
the stable PageRank value as ordinate. It is clear that Angelina Jolie is the most
influential movie star in this network.

5.4 Sensitivity Analysis

To analysis the robustness of our model, we perform a sensitivity analysis on

our approach.In the PageRank algorithm, there is an equation:
Team # 25318 Page 17 of 20

Table 7: PR value and number of Google search results for each star.

Name PR value Google search

results(million)
Angelina Jolie 0.140634 242
Tom Cruise 0.117486 238
Brad Pitt 0.096786 198
Morgan Freeman 0.091811 60.6
Matt Damon 0.087352 85.1
Tom Hanks 0.075816 122
Nicole Kidman 0.073816 86.5
Leonardo DiCaprio 0.070141 149
George Clooney 0.062742 58.1
Nicolas Cage 0.044374 34.8
Sandra Bullock 0.044015 49.5
Joseph Gordon-Levitt 0.032107 19.7
Anne Hathaway 0.032107 85.5
Kirsten Dunst 0.016211 15.9
Tobey Maguire 0.014601 7.44

Figure 9: Authority-Popularity Diagram for the co-star network

X P Rj (k) 1−q
P Ri (k) = q + i = 1, 2, . . . , n
j
L(j) n

where q is a constant, generally be called Damp Coefficient. The meaning of q is:

There is a Possibility, which equal to (1-q), that all of the nodes in the network
have the same PageRank value :1/N . If coefficient q is not used, the ’strong’
nodes will be so strong that other nodes are hard to ’survive’. For the co-star
network, this is also make sense. The PageRank value of a movie star reflects
the authority, it can also be seen as the possibility that a movie wants him to
Team # 25318 Page 18 of 20

star. But there are always some movies do not need the most famous actor, the
movie stars all have the opportunity to be the starring. This situation is reflected
by coefficient q. So value of q can influence the PageRank of the network. q is de-
faulted be 0.85, we set the PageRank for q = 0.85 as standard. With q changes, we
get the corresponding PageRank and analyze the similarity between this PageR-
ank and the standard. The results were shown in Figure 10.

Figure 10: Sensitivity to Damp Coefficient q

The results show that in a certain range, the PageRank change slowly with
the change of q. So we can see that our model has high stability.

5.5 Strengths and Weaknesses

Strengths

• Comprehensiveness: From the perspective of both depth and width, we

determine the importance based on a variety of indexes.

• Adaptability and Practicability: The model we build has good portability,

it is suitable for most network analysis.

• Simplicity and Accuracy: The programs of the model are easy to under-
stand, and the calculations are precise.

• Flexibility: No matter whether the weight is added to the nodes or links,

and whether there are dangling nodes in the network, the model is able to
tackle it.
Team # 25318 Page 19 of 20

Weaknesses

• Data Limitations: The model is applicable to large volumes of data net-

work. Unfortunately, if the data of a network is limited, the error is large.

6. The philosophy of the scale-free network

Barabsi and Albert [1999] pointed out that ER random graph and WS small
world model neglect two important characteristics of actual network.

• Growth: The scale of actual network is growing. For instance, a large

number of papers will be published every month. However, in ER random
graph and WS small world model, the number of nodes in the network is
fixed.
• Preferential attachment: New nodes tend to connect with those nodes
with high connectivity. This phenomenon is also known as ’Rich get rich-
er’ or ’Matthew effect’. We can note that new papers tend to cite those
important and influential papers which have been widely quoted. And
also, actors always try their best to cooperate with those movie stars.

The dynamics in the scale-free network is the philosophy behind the famous
Pareto principle[9] (also known as the 80-20 rule), which states that, for many
events, roughly 80% of the effects come from 20% of the causes.

Shortcut to boost influence

As is proposed above, new nodes tend to connect with important nodes in
the network. However, in the real world, the ’important nodes’ are hard to be ap-
proached by the ordinary one. For instance, for an ordinary network researcher,
it is hard to believe he can have the chance to co-author with a leading figure
in the field of network research as the important ones also tend to cooperate
with other influential people. So, how can we boost our influence as quickly as
possible? The solution we propose by our model is shown as below:

• Determine the ’20%’ in the network according to our Relation Distance

Model or Authority-Popularity Evaluation Model
• Calculate the closeness centrality of the ’80%’ to ’20%’
• Cooperating with the 80% (more approachable relatively) with low close-
ness centrality step by step and finally co-author with the key figure in the
field
Team # 25318 Page 20 of 20

Figure 11: The shortcut to boost influence

References

[1] http://www.imdb.com/

[2] Freeman, L. C. (1979). Centrality in social networks: Conceptual clarifica-

tion. Social Networks, 1(3), 215239.

[3] S. Mahmoodzadeh, J. Shahrabi, M. Pariazar, and M. S. Zaeri, Project Se-

lection by Using Fuzzy AHP and TOPSIS Technique, World Academy of
Science, Engineering and Technology 30 2007, 333-338.

[4] http://en.wikipedia.org/wiki/Centrality

[5] Aaron Clauset, Cosma Rohilla Shalizi and M. E. J. Newman,Power-Law

Distributions in Empirical Data,2007,arXiv: 0706.1062v1.

[6] Barabsi, A-L, and Albert, R. Emergence of scaling in random networks. Sci-
ence, 286:509-512, 1999.

[7] http://www.impactfactorsearch.com

[8] http://www.cs.virginia.edu/oracle/

[9] http://en.wikipedia.org/wiki/Pareto_principle