Sei sulla pagina 1di 11

Kirchhoff Centrality Measure

for Collaboration Network

Vladimir V. Mazalov1(B) and Bulat T. Tsynguev2


1
Institute of Applied Mathematical Research,
Karelian Research Center, Russian Academy of Sciences,
11, Pushkinskaya St., Petrozavodsk, Russia 185910
vmazalov@krc.karelia.ru
2
Transbaikal State University,
30, Aleksandro-Zavodskaya St., Chita, Russia 672039
btsynguev@gmail.com

Abstract. This paper extends the concept of betweenness centrality


based on Kirchhoff’s law for electric circuits from centrality of nodes to
centrality of edges. It is shown that this new measure admits analyti-
cal definition for some classes of networks such as bipartite graphs, with
computation for larger networks. This measure is applied for detecting
community structure within networks. The results of numerical experi-
ments for some examples of networks, in particular, Math-Net.ru (a Web
portal of mathematical publications) are presented, and a comparison
with PageRank is given.

Keywords: Kirchhoff centrality measure · Betweenness centrality ·


Weighted graph · Community structure

1 Introduction

Betweenness centrality is an efficient concept for analysis of social and collabo-


ration networks. A pioneering definition of the betweenness centrality for node
i was given in [9] as the mean ratio of the total number of geodesics (shortest
paths) between nodes s and t and the number of geodesics between s and t that
i lies on, considered for all i, j.
A shortcoming of betweenness centrality is analyzing only the shortest paths,
thereby ignoring the paths being one or two steps longer; at the same time, the
edges on such paths can be important for communication processes in a network.
In order to take such paths into account, Page and Brin developed the well-
known PageRank algorithm for ranking of all pages in the Web’s graph based
on the Markov chain limit theorem. Jackson and Wolinsky [11,12] proposed the
model of communication game in which the utility depends of the structure of
the network. They apply the Myerson value [1,13] to analyse the betweenness of
the nodes in the network. Other allocation rules were proposed in [5,6,16,17].
Brandes and Fleischer [8] and Newman [14] introduced the concept of current

c Springer International Publishing Switzerland 2016
H.T. Nguyen and V. Snasel (Eds.): CSoNet 2016, LNCS 9795, pp. 147–157, 2016.
DOI: 10.1007/978-3-319-42345-6 13
148 V.V. Mazalov and B.T. Tsynguev

flow betweenness centrality (CF-centrality). In [2,8], a graph was treated as an


electrical network with edges being unit resistances. The CF-centrality of an
edge is the amount of current flowing through it with averaging over all source-
destination pairs, when one unit of current is induced at the source and the
destination (sink) is connected to the ground.
In [3], the authors developed this concept further, suggesting to ground all
nodes equally; as a result, averaging runs only over the source nodes, reducing
computational cost. This new approach will be referred to as beta current flow
centrality (βCF-centrality). Moreover, in contrast to the works [2,8,14,15], the
weighted networks were considered in [3], mostly in the context of the βCF-
centrality of nodes. The present paper focuses on the βCF-centrality of edges.
Using this measure, a community detection method is proposed for the networks.
The method is tested on the collaboration network presented by Math-Net.ru, a
Web portal of mathematical publications.

2 Beta Current Flow Centrality Based on Kirchhoff ’s


Law

Recall in brief the concept of βCF-centrality proposed in [3]. Consider a weighted


graph G = (V, E, W ), with V denoting the set of nodes, E the set of edges and
W the matrix of weights, i.e.,
⎛ ⎞
0 w1,2 . . . w1,n
⎜ w2,1 0 . . . w2,n ⎟
⎜ ⎟
W (G) = ⎜ . .. . . .. ⎟ .
⎝ .. . . . ⎠
wn,1 wn,2 . . . 0

Here wi,j  0 is the weight of an edge connecting nodes i and j, n = |V | gives


the number of nodes. Note that wi,j = 0 if nodes i and j are not adjacent. By
assumption, G represents an undirected graph, viz., wi,j = wj,i .
For the weighted graph G, the Laplacian matrix L(G) is defined by
⎛ ⎞
d1 −w1,2 . . . −w1,n
⎜ −w2,1 d2 . . . −w2,n ⎟
⎜ ⎟
L(G) = D(G) − W (G) = ⎜ . .. .. .. ⎟ (1)
⎝ .. . . . ⎠
−wn,1 −wn,2 . . . dn
n
where di = j=1 wi,j yields the sum of weights of the edges adjacent to node i
in the graph G.
Let a graph G be obtained from the graph G by adding single node n + 1
connected to all nodes of the graph G via the links of a constant conductance β.
Kirchhoff Centrality Measure for Collaboration Network 149

Thus, the Laplacian matrix for the modified graph G has the form
⎛ ⎞
d1 + β −w1,2 . . . −w1,n −β
⎜ −w2,1 d2 + β . . . −w2,n −β ⎟
⎜ ⎟
⎜ .. ⎟ .
L(G ) = D(G ) − W (G ) = ⎜ ...
   ..
.
..
.
..
. . ⎟
⎜ ⎟
⎝ −wn,1 −wn,2 . . . dn + β −β ⎠
−β −β . . . −β βn
Suppose that a unit of current flows into node s ∈ V and node n + 1 is grounded.
Let ϕsi be the electric potential at node i when an electric charge is concentrated
at node s. By Kirchhoff’s current law, the vector of all potentials ϕs (G ) =
[ϕs1 , . . . , ϕsn , ϕsn+1 ]T at the nodes of the graph G satisfies the following system
of equations:
L(G )ϕs (G ) = bs , (2)
where bs is the (n + 1)-dimensional vector

 1 i = s,
bs (i) = (3)
0 otherwise.

The Laplacian matrix (1) is singular. The potentials can be determined up to


a constant. Hence, without loss of generality, assume that the potential at node
n + 1 is 0 (a grounded node). Then it follows from (2) that
ϕ̃s (G ) = L̃(G )−1 bs , (4)
where ϕ̃s (G ), L̃(G ) and bs are obtained from (2) by deleting the last row and
column that correspond to node n + 1.
Thus, the vector ϕ̃s (G ) can be considered as the vector of potentials at the
nodes of graph G, i.e.,
ϕ̃s (G) = [L(G) + βI]−1 bs .
According to Ohm’s law, for the link e = (i, j) the let-through current is
xse = |ϕsi − ϕsj | · wi,j . Define the βCF-centrality of edge e as
1
s
CFβ (e) = xe . (5)
n
s∈V

Given that the electric charge is concentrated at node s, the mean value of
the current flowing through node i is
1

xs (i) = (bs (i) + xse ). (6)


2 e:i∈e

And finally, define the beta current flow centrality (βCF-centrality) of node
i in the form
1
s
CFβ (i) = x (i). (7)
n
s∈V
150 V.V. Mazalov and B.T. Tsynguev

3 Unweighted Network of Eleven Nodes


Let us begin with a simple example of an eleven-node network, which elucidates
the weak properties of flow centrality based on geodesics (see Fig. 1). A graph
contains two sets of nodes connected by three nodes 1, 2 and 11. Nodes 2 and
11 have higher values of centrality, as all geodesics between the two sets pass
through these nodes. Hence, node 1 has zero centrality based on geodesics.
Actually, node 1 plays an important role in information distribution. Information
can be transmitted not only directly from one node to another, but also via an
additional node.

Fig. 1. Unweighted network of eleven nodes.

First, compute the βCF-centrality of the nodes in this network with β = 0.5.
This method ranks nodes 2 and 11 as higher ones with centrality value 0.291
and node 1 as third with centrality value 0.147. The other nodes have centrality
values 0.127.
Then, calculate the βCF-centrality of the edges in this graph. The βCF-
centrality of the edge (2, 11) is 0.137, and the centrality of edges (1, 2), (1, 11) is
0.101. The other edges have centrality 0.0647.
In fact, the centrality of nodes 2 and 11 is twice as great as that of node 1. At
the same time, the centrality of node 1 and adjacent edges exceeds the centrality
of the other nodes and edges in the network.

4 Bipartite Unweighted Graph


Consider a bipartite graph G with n nodes divided into two sets V1 and V2 so
that a node from V1 is connected only with a node from V2 , and vice versa (see
Fig. 2). Denote this graph by K|V1 |,|V2 | . Here all links have unit weights.
Kirchhoff Centrality Measure for Collaboration Network 151

4.1 Bipartite Graph K2,n−2


Let V1 = {v1 , v2 }, v  ∈ V2 . Then the Laplacian matrix is

D(G) − W (G) + βI
⎛ ⎞
n−2+β 0 −1 −1 . . . −1
⎜ 0 n − 2 + β −1 −1 . . . −1 ⎟
⎜ ⎟
⎜ −1 −1 2 + β 0 ... 0 ⎟
⎜ ⎟
= ⎜ −1 −1 0 2 + β ... 0 ⎟ .
⎜ ⎟
⎜ .. .. .. .. .. .. ⎟
⎝ . . . . . . ⎠
−1 −1 0 0 ... 2 + β

Its inverse has the form

(D(G) − W (G) + βI)−1


⎛ ⎞
n−2+nβ+β 2 n−2
1 1 ... 1
⎜ n+β−2 n+β−2
n−2+nβ+β 2

⎜ n−2
1 1 ... 1 ⎟
⎜ n+β−2 n+β−2 ⎟
⎜ 2+nβ+β 2 ⎟
1 ⎜ 1 1 2
... 2

= ⎜ 2+β 2+β 2+β ⎟.
β(n + δ) ⎜
⎜ 1 1 2
2+β
2+nβ+β 2
2+β ... 2
2+β


⎜ .. .. .. .. .. ⎟
⎜ .. ⎟
⎝ . . . . . . ⎠
2 2 2+nβ+β 2
1 1 2+β 2+β ... 2+β

For s = v1 , the current is:



s 1 (β + n − 1)(n − 2)
x (v1 ) = 1+ ,
2 (n + β)(n + β − 2)
(n − 2)
xs (v2 ) = ,
2(n + β − 2)(β + n)
1
xs (v  ) = .
2(n + β − 2)
Nodes v1 and v2 are symmetrical; and so, for s = v2 we obtain

(n − 2)
xs (v1 ) = ,
2(n + β − 2)(β + n)

s 1 (β + n − 1)(n − 2)
x (v2 ) = 1+ ,
2 (n + β)(n + β − 2)
1
xs (v  ) = .
2(n + β − 2)
For s = v  ,
β + 2n − 4
xs (v1 ) = xs (v2 ) = ,
2(2 + β)(β + n)
152 V.V. Mazalov and B.T. Tsynguev

s 1 2(β + n − 1)
x (s) = 1+ ,
2 (2 + β)(β + n)
1
xs (v  ) = ,
(2 + β)(β + n)
which yields


1 n−2 (n − 2)(β + 2n − 4)
CFβ (v1 ) = CFβ (v2 ) = 1+ + ,
2n n+β−2 (2 + β)(β + n)

 1 2 2(β + 2n − 4)
CFβ (v ) = 1+ + .
2n n + β − 2 (2 + β)(β + n)

4.2 Bipartite Graph K3,n−3

Let V1 = {v1 , v2 , v3 }, v  ∈ V2 .
In this case, the Laplacian matrix is

D(G) − W (G) + βI
⎛ ⎞
n−3+β 0 0 −1 −1 . . . −1
⎜ 0 n − 3 + β 0 −1 −1 . . . −1 ⎟
⎜ ⎟
⎜ 0 0 n − 3 +β −1 −1 . . . −1 ⎟
⎜ ⎟
⎜ −1 −1 −1 3+β 0 ... 0 ⎟
=⎜ ⎟.
⎜ −1 −1 −1 0 3+β ... 0 ⎟
⎜ ⎟
⎜ .. .. .. .. .. .. . ⎟
⎝ . . . . . . .. ⎠
−1 −1 −1 0 0 ... 3 + β

Its inverse has the form


1
(D(G) − W (G) + βI)−1 =
β(n + β)
⎛ n−3+nβ+δ2 n−3 n−3

n+β−3 n+β−3 n+β−3 1 1 ... 1
⎜ n−3 n−3+nβ+δ 2
n−3 ⎟
⎜ 1 1 ... 1 ⎟
⎜ n+β−3 n+β−3 n+β−3 ⎟
⎜ n−3 n−3 n−3+nβ+δ 2 ⎟
⎜ n+β−3 n+β−3 n+β−3 1 1 ... 1 ⎟
⎜ 3+nβ+δ 2 ⎟
×⎜ ⎟.
3 3
⎜ 1 1 1 3+β 3+β ... 3+β ⎟
⎜ 3 3+nβ+β 2 3 ⎟
⎜ 1 1 1 ... ⎟
⎜ 3+β 3+β 3+β

⎜ .. .. .. .. .. .. .. ⎟
⎝ . . . . . . . ⎠
3 3 3+nβ+β 2
1 1 1 3+β 3+β ... 3+β

By analogy with K2,n−2 , for K3,n−3 it follows that


Kirchhoff Centrality Measure for Collaboration Network 153

CFβ (v1 ) = CFβ (v2 ) = CFβ (v3 )



1 (n − 3)(β + n + 1) (n − 3)(β + 2n − 5)
= 1+ + ,
2n (n + β − 3)(β + n) (3 + β)(β + n)

1 3(β + n + 1) 3(β + 2n − 5)
CFβ (v  ) = 1+ + .
2n (n + β − 3)(β + n) (3 + β)(β + n)

4.3 Bipartite Graph Kr,n−r

Consider the general case where v ∈ V1 , v  ∈ V2 and r = |V 1|, n − r = |V 2|.

Fig. 2. Bipartite graph Kr,n−r .

By induction, for the general graph Kr,n−r ,


1 (n − r)(β + n − 2 + r) (n − r)(β + 2n − 2 − r)
CFβ (v) = 1+ + ,
2n (n + β − r)(β + n) (r + β)(β + n)

 1 r(β + n − 2 + r) r(β + 2n − 2 − r)
CFβ (v ) = 1+ + .
2n (n + β − r)(β + n) (r + β)(β + n)
Observe that all edges have the same βCF-centrality for the bipartite graph
Kr,n−r , i.e.,

1 β+n−2+r β + 2n − 2 − r
CFβ (e) = + .
n (n + β − r)(β + n) (r + β)(β + n)

5 The Results of Computer Experiments


with M ath-N et.ru

Figure 3 shows the subgraph associated with Math-Net.ru, a Web portal of math-
ematical publications. The total amount of the authors on the portal is 78839.
154 V.V. Mazalov and B.T. Tsynguev

Fig. 3. Subgraph of mathematical Web portal M ath-N et.ru.

We will consider only one connected component of this graph with 7606 mathe-
maticians and 10747 publications coauthored by them. The nodes of the graph
describe the authors and the link weights give the number of coauthored publi-
cations. Actually, the publications having more that 6 coauthors are ignored.
For simplicity, all links with the weights smaller than 7 are deleted, see the
result in Fig. 4. Clearly, nodes 40, 34, 56 and 20 represent the centers of “local”
stars and, consequently, must have a high centrality. Note that node 32 also
must have a high centrality, as connecting two separate components.
Table 1 combines the ranking results for the first 11 nodes of the graph using
βCF-centrality (formula (8) with the parameter β = 1), the PageRank algorithm
with the parameter α = 0.85 and electric centrality (CF-betweenness) developed
in [7,8].
As supposed, nodes 40, 34, 56 and 20 have high centrality in all ranking
methods considered. But PageRank assigns a low rank (34) to node 32.
Now, let us detect the community structure of the network adhering to the
approach developed in [10]. The whole idea of this approach lies in the following.
If a network contains communities or groups that are only weakly connected via
a few internal edges, then the edges connecting the communities have a high
βCF centrality. By removing these edges, the groups are separated from each
other, and the underlying community structure of the network is revealed.
Kirchhoff Centrality Measure for Collaboration Network 155

Fig. 4. Main component of subgraph associated with M ath-N et.ru.

Table 1. Ranking results for nodes of M ath-N et.ru

Node Centrality (CFβ ) Node PageRank Node CF-betwenness centrality


40 0.15740 40 0.04438 56 0.54237
34 0.14981 34 0.03285 32 0.53027
20 0.13690 20 0.03210 47 0.48222
47 0.12566 56 0.02774 22 0.41668
56 0.12518 47 0.02088 33 0.41361
26 0.10880 39 0.01874 34 0.39517
30 0.09098 28 0.01824 30 0.39426
9 0.08149 21 0.01695 52 0.37421
33 0.08024 65 0.01632 40 0.36946
32 0.07959 26 0.01552 26 0.35259
22 0.07903 107 0.01424 20 0.34413

First, calculate the βCF -centrality of all edges in the network. Find an edge
with the higher centrality (actually, edge (32,56)) and remove it from the graph.
Next, recalculate the βCF centrality of all edges in the modified network. Again,
find an edge with the higher centrality and remove it from the graph, etc. The
described process is continued until no edges remain.
The results of these computations are presented below.

(32, 56), (9, 30), (47, 52), (20, 75), (22, 26), (34, 119), (128, 132), (9, 11), (4, 5), . . .

After ranking of all edges, all nodes of the network can be divided into the
communities (clusters). Figure 4 shows the resulting community structure of the
collaboration network on the Web portal Math-Net.ru. The graph splits into 7
communities corresponding to different fields of mathematics, namely, coding,
156 V.V. Mazalov and B.T. Tsynguev

discrete mathematics, mathematical physics, functional analysis, algebra and


topology, optimal control, and probability theory (Fig. 5).

Fig. 5. Community structure of the subgraph associated with M ath-N et.ru.

6 Conclusion
This paper has investigated the community structure of networks using a new
concept of betweenness centrality measure. The βCF-centrality measure of the
nodes in a network depending on the parameter β was introduced earlier in [3]
based on electric circuit interpretation. In the present paper, this measure has
been extended to the edges in a network. Moreover, the measure has been applied
for detecting the community structure of networks. The proposed method has
been have tested on the graph of mathematical publications available at the Web
portal Math-Net.ru.

Acknowledgements. This research was supported by the Russian Foundation for


Basic Research (project no. 16-51-55006), the Russian Humanitarian Science Founda-
tion (project no. 15-02-00352) and the Division of Mathematical Sciences of the Russian
Academy of Sciences.

References
1. Aumann, R., Myerson, R.: Endogenous formation of links between players and
coalitions: an application of the Shapley value. In: Roth, A. (ed.) The Shapley
Value, pp. 175–191. Cambridge University Press, Cambridge (1988)
2. Avrachenkov, K., Litvak, N., Medyanikov, V., Sokol, M.: Alpha current flow
betweenness centrality. In: Bonato, A., Mitzenmacher, M., Pralat, P. (eds.) WAW
2013. LNCS, vol. 8305, pp. 106–117. Springer, Heidelberg (2013)
Kirchhoff Centrality Measure for Collaboration Network 157

3. Avrachenkov, K.E., Mazalov, V.V., Tsynguev, B.T.: Beta current flow centrality
for weighted networks. In: Thai, M.T., Nguyen, N.T., Shen, H. (eds.) CSoNet 2015.
LNCS, vol. 9197, pp. 216–227. Springer, Heidelberg (2015)
4. Borgatti, S.P., Everett, M.G., Freeman, L.C.: Ucinet for Windows: Software for
Social Network Analysis. Analytic Technologies, Harvard (2002)
5. Borm, P., Owen, G., Tijs, S.: On the position value for communication situations.
SIAM J. Disc. Math. 5(3), 305–320 (1992)
6. Borm, P., van den Nouweland, A., Tijs, S.: Cooperation and communication restric-
tions: a survey. In: Gilles, R.P., Ruys, P.H.M. (eds.) Imperfections and Behavior
in Economic Organizations. Kluwer Academic Publishers, Boston (1994)
7. Brandes, U.: A faster algorithm for betweenness centrality. J. Math. Sociol. 25,
163–177 (2001)
8. Brandes, U., Fleischer, D.: Centrality measures based on current flow. In: Diek-
ert, V., Durand, B. (eds.) STACS 2005. LNCS, vol. 3404, pp. 533–544. Springer,
Heidelberg (2005)
9. Freeman, L.C.: A set of measures of centrality based on betweenness. Sociometry
40, 35–41 (1977)
10. Girvan, M., Newman, M.E.J.: Community structure in social and biological net-
works. Proc. Natl. Acad. Sci. U.S.A. 99(12), 7821–7826 (2002)
11. Jackson, M.O.: Allocation rules for network games. Games Econ. Behav. 51(1),
128–154 (2005)
12. Jackson, M.O.: Social and Economic Networks. Princeton University Press, Prince-
ton (2008)
13. Mazalov, V.V., Trukhina, L.I.: Generating functions and the Myerson vector in
communication networks. Disc. Math. Appl. 24(5), 295–303 (2014)
14. Newman, M.E.J.: A measure of betweenness centrality based on random walks.
Soc. Netw. 27, 39–54 (2005)
15. Opsahl, T., Agneessens, F., Skvoretz, J.: Node centrality in weighted networks:
generalizing degree and shortest paths. Soc. Netw. 32, 245–251 (2010)
16. Slikker, M., Gilles, R.P., Norde, H., Tijs, S.: Directed networks, allocation proper-
ties and hierarchy formation. Math. Soc. Sci. 49(1), 55–80 (2005)
17. Talman, D., Yamamoto, Y.: Average tree solutions and subcore for acyclic graph
games. J. Oper. Res. Soc. Jpn. 51(3), 187–201 (2008)

Potrebbero piacerti anche