Vous êtes sur la page 1sur 74

Introduction

Centrality
Formulations
Computational Results
Conclusions
References

Detecting Cliques of Maximum and Minimum


Centrality in Information and Social Networks
Chrysafis Vogiatzis
Department of Industrial and Manufacturing Engineering
North Dakota State University
Work funded by the Air Force Research Lab Mathematical Modeling and
Optimization Institute

February 26, 2016


Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Outline
1

Introduction

Centrality

Formulations

Computational Results

Conclusions

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Outline
1

Introduction
Chrysafis
Outline
Graphs

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Who I am
Ph.D. in ISE, Assistant Professor at
IME/NDSU
OR, Mathematical Programming,
Combinatorial Optimization
I come from a small island in Greece
Applications I work on:
Network analysis and optimization
Biological networks
Evacuation and disaster management
Data analytics using combinatorial optimization
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Goals

graphs in decision making


measuring actor importance
applications
group centrality
novel approach/old applications
Figure: The collaboration network

of Paul Erdos.

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Graphs in the modern world

Figure: The city of Le Havre in France

Figure: My Facebook graph

Figure: Myspace trend


Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Why use graphs?


highly connected world
easy to visualize
reveal extra information
natural interpretations

A collection of points together with


lines that connect some subset of the
points."
Wolfram Alpha

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Chrysafis
Outline
Graphs

Graph Theory Preliminaries


G(V , E)
V = {A, B, C, D, E}
E = {AB, AC, AD, AE, CE}
Pij is a geodesic path
d(i, j)

G[S]
C is a clique

N(S) is the neighborhood of S

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Outline
2

Centrality
Preliminaries
Definitions
Complexity

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Preliminaries
Definition
A centrality index is a realvalued function on the nodes of a graph G(V, E).

It is a structural index
It reveals the topographical" importance of a node

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Preliminaries
fundamental concept
measures importance" or influence"
social network analysis, traffic monitoring, data analytics

Social Network Analysis


Traffic Monitoring
Power Systems
Biological Networks

Bavelas [1, 2], Leavitt [7], and Sabidussi [9]


Dolev et al. [4]
Pagani et al. [8]
Wang et al. [10]

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Common Centrality Metrics


Centrality metrics of a node A in a network are commonly
defined as follows:
1
2

Degree centrality: number of nodes with direct links to A


Closeness centrality: Total (maximum) distance to node
A from any other node in the network
Betweenness centrality: number of shortest paths from
any two nodes B, C that pass through A.

For node betweenness, let


gij = the number of geodesics between nodes i and j
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Definitions
Degree

Betweenness

Closeness

others (Katz, alpha, eigenvector,


etc.)

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Centrality examples

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Centrality examples

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Centrality examples

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Extensions
Everett and Borgatti [5, 6]:
extended the definitions to groups (connected and not)
enabled researchers with relative centrality
analyzed influence of specific classes
highest betweenness: prime candidates for information
tracking (e.g., monitoring)
highest degree: prime candidates for information spreading
highest closeness: prime candidate for influence (e.g.,
information control)

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Our contribution: Cliques and Centrality


We consider the problem of detecting cohesive subgraph
centrality.
Why cohesive?
cliques have a wide variety of applications
creation of informal network groups
group centrality is a hard task; the extra properties of
cliques make formulations manageable
a well-connected small clique ends up being more
important than the largest clique
interested in both most and least central cliques
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Centrality extensions
For a set of nodes S V , we have
Degree Centrality:
C d (S) = |N(S)|

(1)

C1c (S) = max d(S, i)


iV \S
X
C2c (S) =
d(S, i)

(2)

Closeness Centrality:

(3)

iV \S

Betweenness Centrality:
C b (S) =

X gS (i, j)
g(i, j)

i,jV \S

Vogiatzis

Clique Centrality

(4)

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Residual Centrality
Node closeness centrality is also defined as the inverse of node
farness, which is the sum of all distances to other nodes on
graph G.
CG (i) =

X 1
dij

(5)

j6=i

Residual centrality is a more novel notion that aims to evaluate


graph centrality after nodes/node structures are removed from
the graph [3].

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Ratio Problem
Residual Centrality:
C r (S) = 1

CG[V \S] (V \ S)
CG (V \ S)

(6)
(7)

Some node examples:

(a) Star
C r (center) = 1,
C r (leaf) = 0

(b) Path

Vogiatzis

(c) Clique
C r (i) = 0

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Residual Centrality: Graph


Example (Clique Removal)

Figure: A graph consisting of 19 nodes.

Figure: The induced subgraph of V \ S, when S is the triangle in the


middle.

C r (S) = 0.5256

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Star Centrality
Consider the superstar" in the figure.
Typically, the blue nodes are more central" (higher degree)
However, in interdependent and biological networks, this
leads to locally" centralized nodes
Now consider the star, centered at the red node

Cstar (i) = max{|N(S)| : S forms an induced star, centered at i}


Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Problem Definition

Definition
CLIQUE CENTRALITY: Given a graph G = (V , E), a centrality
+

metric C() {C d (), C1c (), C2c (), C b (), C b (), C b ()}, a
non-negative number s and an integer k , is there a clique
S V of size k such that C(S) s (or C(S) s)?

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Degree Complexity
Definition
CLIQUE DEGREE
CENTRALITY: Given a graph
G(V , E) and an integer number
k , is there a clique C V such
that the degree of C is at least
k?

(1)

u1

Vogiatzis

(5)

4
2

2
3

...
(1)

u4

(1)

(5)

u3

Figure: Example of the reduction

from G to G.

Clique Centrality

(5)

u4

...
u3

Theorem
CLIQUE DEGREE
CENTRALITY is NPcomplete.

(5) (1)

u1 u2
u2
...
...

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Closeness Complexity
Definition

G0

CLIQUE CLOSENESS
CENTRALITY: Given a graph
G(V , E) and an integer l, is
there a clique C V such that
the distance of every other
node to the clique is at most l?
Theorem
CLIQUE CLOSENESS
CENTRALITY is NP-complete.
Vogiatzis

vu11

vul1
1

vu21
...

u1

vu12

vul1
2

vu22
...

u2
u3

vu13

...
vu14

vul1
4

vu24
...

u4
vu15
u5

vul1
3

vu23

vul1
5

vu25
...

Figure: Reduction for a graph G.

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Preliminaries
Definitions
Complexity

Betweenness Complexity
Definition
CLIQUE BETWEENNESS
CENTRALITY: Given a graph
G(V , E) and an integer number
k , is there a clique C V such
that the number of shortest
paths that go through at least
one of the nodes v C is at
least equal to k ?
Theorem
CLIQUE BETWEENNESS
CENTRALITY is NPcomplete.
Vogiatzis

G0

vu11

vul1
1

vu21
...

u1

vu12

vul1
2

vu22
...

u2
u3

vu13

...
vu14

vu24

vu15

vu25

vul1
4
...

u4
u5

vul1
3

vu23

vul1
5
...

Figure: Reduction for a graph G.


Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Outline
3

Formulations
Degree Centrality
Closeness Centrality
Betweenness Centrality

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Overview

detect cliques of max/min centrality


cliques are assumed to be of equal size k
degree/closeness/betweenness
betweenness
probabilistic
optimistic
pessimistic

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Max Degree
Let

xi =

if i V is in the clique C
otherwise

1,
0,

and
yi = 1, iff (i, j) E, i V \ C, j C.
Then,

(MAX-C d ) : max

yi

(8)

iV

s.t.
xi + yi 1,
X
yi
xj ,

i V

(9)

i V

(10)

(i, j)
/E

(11)

j:(i,j)E

xi + xj 1,
X
xi = k ,

(12)

iV

xi , yi {0, 1},

i V
Vogiatzis

(13)
Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Min Degree
(MIN-C d ) : min

|V |
X

yi

(14)

i=1

s.t.
(10), (11), (12),
1 X
xj ,
xi + yi
i

i V

(15)

i V

(16)

j:(i,j)E

xi , yi {0, 1},

i is a sufficiently large constant


for a tighter LP it can be set to i = min(|N(i)|, k )
Theorem
When yi is allowed to be continuous, resulting in formulations
RMAX-C d and RMIN-C d , there always exists an optimal solution
(x , y ) that is binary.
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Closeness
C1c (S) = max d(S, i) or Bottleneck version
iV\S

S = {4, 5} has a bottleneck closeness of 2.

C2c (S) =

d(S, i) or Total version

iV\S

S = {2, 3, 4} has a minimum total closeness of 8.

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Closeness
(k )

xi

(0)
xi

: a new (recursively defined) variable for every node i V


xi

(`)
xi

= 1: diS `
P (`) P (`1)
v` = 1 iff
xi
xi
1
iV

(`)

xi

iV

(`1)

xj

(`1)

+ xi

i V , ` = 1, . . . , diam(G)

j:(i,j)E

(17)

(`)

xi

X (`1)
1 (`1)
, i V , ` = 1, . . . , diam(G)
xi
+
xj
i
j:(i,j)E

(18)
i can be set to |N(i)| + 1
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Closeness
diam(G)

(MIN-C1c ) : min

v`

(19)

`=1

s.t.
(11), (12), (17), (18)
X (`)
(`1)
(xj xj
),
v`

1 ` diam(G)

jV

(20)
v`

1 X (`)
(`1)
(xj xj
),
|V |

1 l diam(G)

jV

(21)
(`)
xi , v`

{0, 1},

i V , 0 ` diam(G)

diam(G)

(MIN-C2c ) : min

`=1

!
`

(`)

xi

iV

(`1)

xi

(22)

iV

s.t.
(11), (12), (17), (18)
(`)

xi

{0, 1}

i V , 0 ` diam(G)

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness
Probabilistic Betweenness
2

S = {2} has a probabilistic betweenness of


(1, 4).

1
2

when considering pair

Optimistic Betweenness
2

S = {2} has an optimistic betweenness of 1 when considering pair


(1, 4).

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness
Pessimistic Betweenness
2

S = {2} has a pessimistic betweenness of 0 when considering pair


(1, 4).

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness

Figure: An illustrative example that computing individual node


b
betweenness centralities is notj enough
k to compute C (S).
j Onkboth

a (i,j)
= 0 and C b (b) =
figures: S = {a, b}, C b (a) = gg(i,j)
j
k
gS (i,j)
b
However, C (S) = g(i,j) = 1 on the left figure, and
j
k

S (i,j)
C b (S) = gg(i,j)
= 0 on the right figure.

Vogiatzis

gb (i,j)
g(i,j)

Clique Centrality

= 0.

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness
Let gk (i, j) be the number of geodesics Pij that pass through
node k V . Also, let ge (i, j) be the number of geodesics Pij
that use edge e E.
Proposition
Let S V be a clique in a graph G = (V , E) with the set of
edges Es = (S S) E. Then, for any pair of nodes i, j V \ S:
X
X
gS (i, j) =
gt (i, j)
ge (i, j).
(23)
tS

Vogiatzis

eEs

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness Probabilistic

need to compute gt (i, j) and ge (i, j) for all i, j, t V and


eE
let ue (e E) be a binary variable such that ue = 1 iff
e Es
to calculate gtij , we use a modified version of the classical
all-pairs shortest path algorithm of Floyd-Warshall
ge (i, j) is a byproduct of the previous calculations

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness Probabilistic
P
P ij
ij
g u
g x
X tV t t eE e e

(MAX-C b ) : max

g ij

(24)

i,jV

s.t.
(11), (12),
ue xi , ue xj , ue xi + xj 1,

e = (i, j) E

(25)

i V , e E

(26)

xi , ue {0, 1},
where constraints (25) ensure that ue = 1 if and only if e ES .

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness Optimistic
Observe that
+

C b (S) =

X
i,jV \S


max
tS

gt (i, j)
g(i, j)


(27)

Now, let

zij =

1,

0,

iff there exists a shortest path between nodes i and j


which passes through at least one node in the clique S
otherwise.

and
ij

ft =


gt (i, j)
, i, j, t V
g(i, j)

Vogiatzis

Clique Centrality

(28)

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness Optimistic
+

(MAX-C b ) : max

zij

(29)

i,jV

s.t.
(11), (12),
1 X ij
zij
ft xt xi xj ,
|V |
lV
X ij
zij
ft xt

i, j V

(30)

i, j V

(31)

i, j V

(32)

tV

xi , zij {0, 1},

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Degree Centrality
Closeness Centrality
Betweenness Centrality

Betweenness Pessimistic
Let

tij =

1,
0,

iff all geodesics between nodes i, j pass through a clique S


otherwise.

(MAX-C b ) : max

tij

(33)

i,jV

s.t.
(11), (12), (25)
X ij
X ij
g ij tij
gt xt
ge ue ,
tV

i, j V ,

eE

(34)
tij + xi 1, tij + xj 1,
xi , tij , ue {0, 1},

Vogiatzis

i, j V ,
(35)
i, j V , e E
(36)
Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Outline
4

Computational Results
Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

(a) Node

(b) Triangle

(c) 4-clique

(d) 5-clique

(a) 6-clique

(b) 7-clique

(c) 8-clique

(d) Max Clique

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Les Miserables

(a) Node

(b) Triangle

(c) 4-clique

(d) 5-clique

(a) 6-clique

(b) 7-clique

(c) 8-clique

(d) Max Clique

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Numerical Results
Table: Maximum and minimum clique centralities in book graphs.
D EGREE (%)
|S|

C d ()
Max
Min

1
2
3
4
5
6
7
8
9
10
11

72.60 1.37
79.17 1.38
83.10 1.41
85.71 7.14
88.41 7.79
88.24 4.42
88.06 4.48
80.30 4.55
80.00 15.38
79.69 17.19
73.02 73.02

1
2
3
4
5
6
7
8
9
10

45.57 1.27
61.55 1.28
67.53 5.19
69.74 3.95
66.67 2.67
64.86 2.70
64.38 4.11
34.72 9.72
33.80 11.27
32.86 21.43

C LOSENESS
C1c ()
MinMax

C2c ()
MinMax

B ETWEENNESS (%)
C b ()
Max Min

C b ()
Max Min

C b ()
Max Min

huck: |V | = 74, |E| = 303, max clique size = 11


3
3
3
3
3
3
3
3
3
4
4

6
6
6
5
5
5
5
5
5
5
4

1.384.89
1.314.90
1.283.96
1.262.27
1.232.27
1.242.29
1.242.31
1.322.33
1.322.09
1.332.07
1.401.40

56.94 1.39
69.79 1.44
78.64 2.98
81.49 3.47
84.93 4.83
87.34 6.49
87.34 9.72
78.5312.13
78.5615.20
78.5620.14
70.1244.97

33.46 0.00
38.63 0.00
41.58 0.00
43.07 0.00
44.15 0.00
44.94 0.00
44.94 0.00
41.70 0.00
41.71 0.05
41.71 1.88
38.7413.37

77.67 2.70
83.89 5.37
86.30 8.00
88.2610.59
89.5613.14
90.4115.66
90.4118.14
87.2320.58
87.2624.86
87.2628.38
83.1983.19

jean: |V | = 80, |E| = 257, max clique size = 10


3
3
3
3
3
3
3
4
4
4

6
5
5
5
5
5
5
4
4
4

1.614.22
1.423.26
1.363.06
1.343.09
1.373.12
1.393.15
1.402.78
1.862.26
1.872.25
1.892.13

51.30 1.28
60.51 1.35
68.00 2.68
73.16 3.14
73.54 4.23
73.54 4.75
73.99 6.55
40.44 7.72
40.57 9.28
40.6612.22

26.36
30.83
34.33
36.03
35.31
34.81
34.90
19.15
19.19
19.22

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.02
0.04
0.58

71.65 2.5
80.63 4.97
86.30 7.41
88.96 9.81
87.3412.18
84.2114.53
84.3016.84
49.7219.98
49.8125.24
49.8736.57

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Numerical Results
Table: Maximum and minimum clique centralities in book graphs
(contd).
D EGREE (%)
|S|

C d ()
Max
Min

C LOSENESS
C1c ()
MinMax

C2c ()
MinMax

B ETWEENNESS (%)
C b ()
Max Min

C b ()
Max Min

C b ()
Max Min

david: |V | = 87, |E| = 406, max clique size = 11


1
2
3
4
5
6
7
8
9
10
11

95.35
100
100
100
100
100
100
100
100
100
100

1.16
1.18
4.76
3.61
6.10
4.93
15.00
26.58
32.05
64.94
100

1
2
3
4
5
6
7
8
9
10
11

51.82
72.79
81.48
87.31
91.73
93.94
94.66
88.46
88.37
76.56
76.38

0.73
0.74
2.96
2.99
3.76
5.30
28.24
42.31
51.16
59.38
76.38

2
1
1
1
1
1
1
1
1
1
1

3
3
3
3
3
2
2
2
2
2
1

1.052.53
12.55
12.55
12.57
11.94
11.95
11.85
11.73
11.68
11.35
1 1

57.58 1.18
64.50 1.20
69.07 2.45
73.56 3.60
79.34 4.56
82.01 6.95
83.40 8.72
83.6710.44
84.1812.17
84.8719.08
84.9029.23

38.13 0.00
42.63 0.00
44.83 0.00
46.65 0.00
48.35 0.00
49.16 0.00
49.60 0.00
49.87 0.00
50.06 0.00
50.23 0.89
50.2515.51

88.24 2.30
92.19 4.57
92.97 6.82
93.69 9.04
94.3311.23
94.7613.39
95.1115.75
95.3518.47
95.5421.72
95.7038.60
95.7595.72

anna: |V | = 138, |E| = 493, max clique size = 11


3
3
2
2
3
3
3
3
3
3
3

5
5
4
4
4
4
4
4
4
4
3

1.534.24
1.283.26
1.192.69
1.132.69
1.092.47
1.072.33
1.061.74
1.121.60
1.121.50
1.251.42
1.251.25

28.01 0.74
41.72 0.75
53.72 1.56
68.62 2.33
75.37 3.13
82.12 4.04
86.91 7.98
77.7610.43
78.4215.77
68.3423.39
68.3438.15

13.66 0.00
19.77 0.00
24.91 0.00
29.31 0.00
31.66 0.00
33.18 0.00
33.90 0.88
31.74 1.26
31.82 3.04
28.37 7.84
28.3721.81

55.33 1.45
73.91 2.89
83.29 4.32
89.87 5.73
94.20 7.15
96.42 8.71
97.2316.76
92.8221.56
92.8929.90
84.1745.65
84.1784.17

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Social Network Benchmarks


Table: Maximum and minimum clique centralities in real-life social and
power grid network instances.
D EGREE (%)
|S|

C d ()
Max
Min

C LOSENESS
C1c ()
MinMax

C2c ()
MinMax

B ETWEENNESS (%)
C b ()
Max Min

C b ()
Max Min

C b ()
Max Min

karate: |V | = 34, |E| = 78, max clique size = 5


1
2
3
4
5

51.51 3.12
62.50 9.38
64.52 9.68
63.33 40.00
62.07 58.62

1
2
3
4
5
6

36.07 1.64
51.67 3.33
55.93 6.78
48.27 12.07
47.37 15.79
42.86 30.36

1
2
3
4
5

19.67 1.64
30.00 3.33
33.90 5.08
31.03 13.79
28.07 21.05

1
2
3
4

7.69
10.34
10.43
6.14

3
2
2
2
2

5
5
4
4
3

1.763.52
1.382.66
1.352.68
1.371.80
1.381.45

39.39 3.13
51.69 4.36
59.71 7.96
67.2010.62
67.7415.24

16.970.00
21.160.00
23.370.00
24.791.38
24.952.59

54.19 5.88
66.1311.59
68.9817.11
70.9439.48
79.7669.16

krebs: |V | = 62, |E| = 153, max clique size = 6


3
2
2
3
3
3

5
5
5
4
4
4

1.704.00
1.553.25
1.513.08
1.592.48
1.602.37
1.642.18

49.34 1.67
61.13 2.00
68.38 3.95
72.61 6.20
73.0811.34
73.0815.56

30.360.00
35.280.00
37.890.00
38.910.21
38.961.03
38.683.32

72.71 3.23
77.63 6.40
81.17 9.52
81.8614.67
81.7623.51
80.1232.05

dolphins: |V | = 62, |E| = 159, max clique size = 5


5
4
4
5
6

8
8
7
7
7

2.395.61
2.134.68
2.023.76
2.173.38
2.723.25

16.71 1.67
33.00 2.06
36.70 4.55
33.00 9.11
35.2716.58

9.080.00
13.790.00
15.500.00
13.700.88
11.955.95

37.54 3.23
49.34 6.40
55.6310.34
51.0816.23
37.1227.74

ieeebus: |V | = 118, |E| = 179, max clique size = 4


0.85
0.86
2.61
6.14

7 14
7 14
8 12
9 9

4.258.85
3.978.84
3.987.03
5.365.36

27.15
44.94
37.17
8.62

0.86
1.19
2.26
5.05

14.160.00
21.040.00
18.080.00
4.121.03

40.56 1.69
54.21 3.38
49.79 5.85
10.5510.55

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Social Network Benchmarks (cont.)


Table: Computational times (in seconds) for solving the maximum and
minimum clique centrality problems in real-life social and power grid
network instances from Table 3.
D EGREE (%)
|S|

C d ()
Max
Min

C LOSENESS
C1c ()
Min Max

B ETWEENNESS (%)

C2c ()
Min Max

C b ()
Max Min

C b ()
Max Min

C b ()
Max Min

karate: |V | = 34, |E| = 78, max clique size = 5


1
2
3
4
5

0.02
0.03
0.65
0.64
0.60

0.03
0.04
0.11
0.10
0.09

1
2
3
4
5
6

0.05
0.16
0.15
0.14
0.05
0.04

0.03
0.16
0.16
0.09
0.05
0.04

1
2
3
4
5

0.05
0.14
0.31
0.33
0.06

0.03
0.07
0.12
0.26
0.05

1
2
3
4

0.24
0.92
0.58
0.38

0.27
1.03
0.59
0.51

0.11
0.25
0.44
0.33
0.30

0.04
0.10
0.43
0.27
0.26

0.06
0.15
0.16
0.13
0.13

0.05
0.09
0.10
0.10
0.10

0.10
0.61
0.58
0.11
0.06

0.12
0.78
0.04
0.10
0.09

0.010.01
0.040.02
0.030.02
0.020.01
0.020.01

0.27
1.13
1.24
0.98
0.45

0.03
0.58
1.01
0.86
0.37

0.45
0.67
1.55
2.20
0.68
0.83

0.08
1.30
2.17
1.70
0.47
0.31

1.09
1.92
4.74
19.42
17.53

0.06
1.10
1.88
2.24
0.66

krebs: |V | = 62, |E| = 153, max clique size = 6


0.16
0.17
0.24
0.13
0.08
0.04

0.17
0.20
0.25
0.13
0.08
0.04

0.12
0.25
0.41
0.13
0.08
0.07

0.10
0.30
0.52
0.16
0.08
0.06

0.27
1.44
1.21
1.73
0.50
0.25

0.05
5.28
0.37
0.60
0.27
0.12

0.030.01
0.360.02
0.280.02
0.100.01
0.100.01
0.080.01

dolphins: |V | = 62, |E| = 159, max clique size = 5


0.23
1.87
0.21
0.41
0.12

0.22
0.40
0.23
0.26
0.11

0.27
0.60
0.77
0.94
0.13

0.17
0.21
0.26
0.34
0.13

0.42
4.07
8.79
7.60
3.42

0.03
2.47
6.88
0.79
0.17

0.030.01
0.460.02
0.470.02
0.480.01
0.070.01

ieeebus: |V | = 118, |E| = 179, max clique size = 4


0.31 0.32
22.7221.95
22.8314.47
10.86 5.21

10.9111.28
26.9517.17
13.04 7.53
4.78 4.35

5.10 0.35
28.6431.04
30.02 2.98
14.53 1.88

0.190.01
1.740.03
0.560.02
0.420.01

5.91 0.56
354.70343.09
272.58234.28
171.83100.29

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

An application in finance
A well-known problem: predict stock market movements.

We created market graphs based on:


standard Pearson correlation
sign correlation
cosine similarity
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Pearson correlation
Figure: Stock market graph with
= 0.6
Ri (t) = ln

cij =

Pi (t)
Pi (t 1)

E(Ri Rj ) E(Ri )E(Rj )


p
Var (Ri )Var (Rj )

If two stocks, i and j, have cij , then i and j are connected with an edge in the stock
market graph.

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Sign correlation
If two stocks, i and j, have the same trend for at least % of the
time, then i and j are connected with an edge in the stock
market graph.
Figure: Stock market graph with = 60

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Cosine similarity
Figure: Stock market graph with
= 0.6
similarity =

Pi Pj
||Pi || ||Pj ||

If two stocks, i and j, have the similarity at least , then i and j are connected with an
edge in the stock market graph.
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Stock Degree Distributions


Pearson

0.06

0.05

0.04

0.03

0.02

0.01

0
0.2

Sign

0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.05

0.04

0.03

0.02

0.01

0
0.2

Cosine

0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.06

0.05

0.04

0.03

0.02

0.01

0
0.2

0.1

0.1

0.2

Vogiatzis

0.3

0.4

0.5

0.6

0.7

0.8

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Detecting central stocks


Pearson
Size=3
Size=4
Sign
Size=3
Size=4
Cosine
Size=3
Size=4

CGG VERITAS, SAINT GOBAIN,


FRANCE TELECOM
NATIXIS, BNP PARIBAS, METROPOLE TV, VINCI
VALLOUREC, IMERYS, FAURECIA
VALLOUREC, IMERYS, FAURECIA, ALSTOM
EURAZEO, WENDEL, VINCI
EURAZEO, WENDEL, VINCI, GECINA NOMINATIF

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Les Miserables
Other Social Networks
Stock Market Analysis
Computational Results

Predicting CAC 40

Pearson: prediction of 86.7% (size=4) and 83.1% (size=3)


Sign: prediction of 88.5% (size=4) and 85.2% (size=3)
Cosine: prediction of 85.8% (size=4) and 83.3% (size=3)
Average prediction rate for stock combinations of size=3, 4
was less than 50%
With merely 4 stocks, we were able to detect the trend of index
CAC40!

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Concluding Remarks
Future directions

Outline
5

Conclusions
Concluding Remarks
Future directions

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Concluding Remarks
Future directions

Conclusions

Presented group extensions to centrality


Clique restriction makes the formulation manageable
Can be viewed as a critical clique" problem
Max Cliques are not always critical"
Obtained a cohesive subset to study the whole stock
market graph

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Concluding Remarks
Future directions

Future Work
Future Directions
Formulate residual centrality for clusters
Detection of special structure" outliers in social networks
Ratio problems on graphs
Centrality in interdependent networks
Special Thanks To
Alexander Veremyev
Eduardo Pasiliao
Jose L. Walteros
Panos M. Pardalos
Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

Concluding Remarks
Future directions

Thank you

Questions?

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

References I
Alex Bavelas. A mathematical model for group
structures. In: Human organization 7.3 (1948),
pp. 1630.
Alex Bavelas. Communication patterns in task-oriented
groups. In: The Journal of the Acoustical Society of
America 22.6 (1950), pp. 725730.
Chavdar Dangalchev. Residual closeness in networks.
In: Physica A: Statistical Mechanics and its Applications
365.2 (2006), pp. 556564.

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

References II
Shlomi Dolev et al. Incremental deployment of network
monitors based on group betweenness centrality. In:
Information Processing Letters 109.20 (2009),
pp. 11721176.
Martin G Everett and Stephen P Borgatti. Extending
centrality. In: Models and methods in social network
analysis 35.1 (2005), pp. 5776.
Martin G Everett and Stephen P Borgatti. The centrality
of groups and classes. In: The Journal of mathematical
sociology 23.3 (1999), pp. 181201.

Vogiatzis

Clique Centrality

Introduction
Centrality
Formulations
Computational Results
Conclusions
References

References III
Harold J Leavitt. Some effects of certain communication
patterns on group performance. In: The Journal of
Abnormal and Social Psychology 46.1 (1951), p. 38.
Giuliano Andrea Pagani and Marco Aiello. The power
grid as a complex network: a survey. In: Physica A:
Statistical Mechanics and its Applications (2013).
Gert Sabidussi. The centrality index of a graph. In:
Psychometrika 31.4 (1966), pp. 581603.
Jianxin Wang, Wei Peng, and Fang-Xiang Wu.
Computational approaches to predicting essential
proteins: A survey. In: PROTEOMICS-Clinical
Applications 7.1-2 (2013), pp. 181192.
Vogiatzis

Clique Centrality

Vous aimerez peut-être aussi