Discriminating Power of Centrality Measures(working paper)

# Discriminating Power of Centrality Measures (working paper)

M. Puck Rombach111Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute, University of Oxford, rombach@maths.ox.ac.uk    Mason A. Porter33footnotemark: 3
###### Abstract

The calculation of centrality measures is common practice in the study of networks, as they attempt to quantify the importance of individual vertices, edges, or other components. Different centralities attempt to measure importance in different ways. In this paper, we examine a conjecture posed by E. Estrada regarding the ability of several measures to distinguish the vertices of networks. Estrada conjectured [12, 9] that if all vertices of a graph have the same subgraph centrality, then all vertices must also have the same degree, eigenvector, closeness, and betweenness centralities. We provide a counterexample for the latter two centrality measures and propose a revised conjecture.

22footnotetext: Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute and CABDyN Complexity Centre, University of Oxford, porterm@maths.ox.ac.uk

## 1 Introduction

Many problems in the study of networks focus on flows along the edges of a network . Although flows such as the transportation of physical objects or the spread of disease are conservative, the flow of information tends to diminish as it moves through a network, and the influence of one vertex on another decreases with increasing network distance between them .

Such considerations have sparked an interest in network communicability, in which a perturbation of one vertex is felt by the other vertices with differing intensities . These intensities depend on all paths between a pair of vertices, though longer paths have smaller influence. This idea is expressed in a general communicability function

 ci=∞∑k=0n∑j=1ck(Ak)ij, (1)

where is the adjacency matrix—whose entries are 1 if vertices and are connected to each other and 0 if they are not—and is the total number of vertices in the network. The networks that we consider in the present paper are unweighted, undirected graphs with no self-edges and no multi-edges. A special version of communicability is the Katz centrality , for which (also see Bonacich power centrality ). Reference  gives an extensive review of network communicability.

Note that one can compute the self-communicability of vertices:

 cii=∞∑k=0ck(Ak)ii. (2)

From a sociological perspective, one can construe as the “karma” of vertex , as it measures its ability to receive its own influence via the graph.

Subgraph centrality, which was introduced in Ref. , is an example of a communicability measure and is defined as

 sci=∞∑k=0(Ak)iik!. (3)

It is a special case of self-communicability in which the weighting function is (which has computational advantages). The sum in (3) is bounded by , so it converges. Subgraph centrality can be computed from the spectrum of the adjacency matrix :

 sci=n∑j=1(vij)2eλj,

where is the th eigenvalue of and is the th component of its associated eigenvector. We will denote the vector containing all vertex subgraph centralities as and use similar notation for other centrality measures.

## 2 Centrality Measures and Their Power To Discriminate

There exist many methods to measure the centrality, or importance, of vertices or other components in a graph . We consider measures of vertex centrality. Different methods for measuring vertex centrality are appropriate for different types of networks, dynamical processes on networks, and applications.

In this paper, we consider subgraph centrality and four of the traditional notions of centrality: degree, closeness, betweenness, and eigenvector centralities. The degree centrality (i.e., degree) of vertex is

 di=n∑j=1Aij (4)

and is simply the number of edges incident on the vertex.

The eigenvector centrality of a vertex is the value of the th entry of the leading eigenvector of the adjacency matrix :

 eci=1λ0∑j∈Γ(i)ecj. (5)

One way to think of the eigenvector centrality of a vertex is as the sum over of the number of distinct paths of length that start at vertex normalized by the total number of paths of length in a graph.

Closeness centrality has been defined in multiple ways in the literature, but it is always a function of the mean length of the shortest paths from a vertex to each of the other vertices in a graph [23, 8]. We use the definition [13, 14]

 cci=n−1∑j∈V(G)P(i,j), (6)

where is the length of a shortest (i.e. geodesic) path between and . The definition (6) is the same as the one used in Ref. .

The geodesic betweenness centrality of a vertex measures how often it lies on shortest paths between pairs of nodes. It is defined as 

 bci=∑j,k∈V(G)∖iP∗jk(i)P∗jk, (7)

where is the number of distinct shortest paths running between vertices and , and is the number of such paths that include vertex .

It was noted in Ref.  that centrality measures such as degree, closeness, betweenness, and eigenvector centralities are unable on certain graphs to distinguish between any of its vertices—even for graphs that are not vertex-transitive. By definition, a graph is “vertex-transitive” if its vertices are indistinguishable: in other words, for all pairs of vertices and , there exists an automorphism that maps to .

In Fig. 1, we show the Frucht graph, which is a cubic (i.e., 3-regular) graph on 12 vertices that has no nontrivial symmetries (i.e., it has no automorphisms other than the identity). This is the smallest regular graph  with this property. Because degree centrality and eigenvector centrality can be described purely in terms of the degrees of vertices and those of their neighbors, these two measures are unable to distinguish between the vertices of the Frucht graph.

In Fig. 2, we show a graph on 6 vertices that is clearly not vertex-transitive (as there exist vertices of degrees 3 and 4) and for which all vertices have the same value for betweenness centrality . Such graphs are called betweenness-self-centric (or are said to have a betweenness centralization equal to ), and Ref.  includes several examples of such graphs.

In Fig. 3, we show a graph on 8 vertices that is not vertex-transitive (three vertices form part of a triangle, but the other five do not) and for which closeness centrality is unable to distinguish the vertices from each another . In this example, degree centrality and eigenvector centrality are also unable to distinguish vertices from each other. Figure 1: The Frucht graph, which is an example of a asymmetric graph for which degree and eigenvector centrality are unable to distinguish the vertices from each other. Figure 2: A graph on 6 vertices that is not vertex-transitive and on which betweenness centrality is unable to distinguish the vertices from each other. Figure 3: A graph on 8 vertices that is not vertex-transitive and on which closeness, degree, and eigenvector centralities are unable to distinguish the vertices from each other.

E. Estrada posed the following two conjectures concerning the discriminating power of subgraph centrality.

 .

###### Conjecture 3.2.

 is walk-regular.

The quantity denotes the variance of , where we have interpreted the vector of centrality values as a set. When , all vertices have the same centrality value.

In this section, we show that Conjecture 3.1, which was originally posed in Ref. , is at least partially false by giving a counterexample for betweenness centrality and closeness centrality.

A graph is called walk-regular if, for every , the number of closed walks in of length that start at vertex is the same for all . Because depends only on the number of closed walks (i.e., walks that start and end at the same vertex) of length starting at , we obtain the following corollary.

###### Corollary 3.3.

 is walk-regular  .

A graph is walk-regular if and only if for all and for all vertices  , where denotes the probability that an ordinary random walk starting at is back at its initial point after steps. All vertex-transitive graphs are walk-regular, but the converse is not true .

A distance-regular graph is a -regular graph with diameter and a set , with and , such that all vertices with distance (i.e., shortest path length) between them satisfy the following property: the number of vertices in that are adjacent to is , and the number of vertices in that are adjacent to is  [5, 21]. All distance-regular graphs are walk-regular, but a distance-regular graph need not be vertex-transitive .

There exist walk-regular graphs that are neither vertex-transitive nor distance-regular . We show an example of such a graph in Fig. 4.***Note that the paper  has a mistake in the plot of this graph. We plot the correct graph in Fig. 4, and one can also find the correct graph in the erratum appended to the paper at http://cs.anu.edu.au/people/bdm/papers/WalkRegular.jpg. This graph on 12 vertices, which we call the Godsil-McKay graph, gives a counterexample to Conjecture 3.1. Its centrality vectors are

 scGM ≈6.7035×[1,1,1,1,1,1,1,1,1,1,1,1], bcGM =[7,8,7,8,7,7,7,7,8,7,8,7], ccGM =[118,119,118,119,118,118,118,118,119,118,119,118].

In Fig. 4, we use gray diamonds to designate the vertices with a betweenness centrality of 8 and a closeness centrality of ; we use blue disks to designate vertices with a betweenness centrality of 7 and a closeness centrality of . Figure 5 shows three more examples of graphs that are walk-regular, but neither distance-regular nor vertex-transitive. Betweenness centrality can distinguish all three of these examples, whereas closeness centrality can distinguish only two of them. Figure 4: The Godsil-McKay graph, which is walk-regular but neither vertex-transitive nor distance-regular . Figure 5: Three more examples of graphs that are walk-regular, but neither distance-regular nor vertex-transitive: (a) a 5-regular graph on 12 vertices, (b) a 4-regular graph on 15 vertices and (c) a 3-regular graph on 20 vertices. Betweenness centrality can distinguish (a), (b), and (c). Closeness centrality can distinguish (b) and (c). Figure 6: Two examples of graphs that are distance-regular (and hence also walk-regular) but not vertex-transitive: (a) the Tutte 12-cage  and (b) the first Chang graph .

In Fig. 6, we give two examples of graphs that are distance-regular (and hence walk-regular) but not vertex-transitive. The examples are the Tutte 12-cage [24, 2, 22] and one of the Chang graphs [7, 25]. The vertices of these graphs cannot be distinguished by degree, eigenvector, closeness, betweenness, or subgraph centralities. The Tutte-12 cage and the Chang graphs are the only known distance-regular graphs that are not also vertex-transitive.See, e.g., the discussion at http://mathoverflow.net/questions/106589/is-every-distance-regular-graph-vertex-transitive. Additionally, it has been proven that betweenness centrality cannot distinguish distance-regular graphs.

###### Lemma 3.4.

 is distance-regular  .

In fact, this is also true for closeness centrality.

###### Lemma 3.5.

is distance-regular  .

###### Proof.

Let , where and is the distance between vertex and . This is a generalization of the neighborhood . Because we are only considering connected graphs, every is associated with a finite positive integer . Therefore, the sets and are pairwise disjoint for any and (with ), and . Given node , we consider the subgraph of that consists of the union of and and all edges between the two sets. We also recall that a biregular graph is a bipartite graph in which all vertices in the same partite set have the same degree. Let be an arbitrary vertex in and let be an arbitrary vertex in . From the definition of a distance-regular graph, we know that

 |Γ(j)∩Γr+1(i)| =br, |Γ(k)∩Γr(i)| =cr+1. (8)

Therefore, is a biregular graph. Additionally, for all , so

 |Γr(i)|=br−1br−2…b0crcr−1…c1 (9)

for each . Because

 cci=n−1∑j∈V(G)P(i,j)=n−1∑dr=1r|Γr(i)| (10)

depends only on (where ), which is independent of the node , it follows that for distance-regular graphs. ∎

The converse of Lemma 3.5 is not true (see the counterexample in Fig. 3).

## 4 Conclusion

Based on the results in Section 3, we propose the following conjecture, which we call the “Modified Estrada Conjecture”.

###### Conjecture 4.1.

is walk-regular, and
is distance-regular.

Note that the parts of the statements in Conjecture 4.1 are known to be true because of Corollary 3.3 and Lemmas 3.4 and 3.5 (and the fact that all distance-regular graphs are walk-regular).

## Acknowledgements

This work was funded by the James S. McDonnell Foundation (#220020177). We thank James Fowler for helpful discussions and Brendan McKay for sending us the examples of walk-regular, non-distance-regular, non-vertex transitive graphs in Fig. 5. We thank Ernesto Estrada for sending us a copy of Ref.  prior to its publication.

## References

•  G. Baron and W. Imrich. Asymmetrische reguläre graphen. Acta Mathematica Academiae Scientiarum Hungaricae, 20:135–142, 2009.
•  C. T. Benson. Minimal regular graphs of girth 8 and 12. Canadian Journal of Mathematics, 18:1091–1094, 1966.
•  P. Bonacich. Factoring and weighting approaches to status scores and clique identification. Journal of Mathematical Sociology, 2:113–120, 1972.
•  P. Bonacich. Power and centrality: A family of measures. American Journal of Sociology, 92(5):1170–1182, 1987.
•  A E Brouwer, A M Cohen, and A Neumaier. Distance-Regular Graphs. Springer, 1989.
•  P. J. Cameron. Random strongly regular graphs? Discrete Mathematics, 273(1):103–114, 2003.
•  L.-C. Chang. The uniqueness and non-uniqueness of the triangular association scheme. Scientific Record of the Peking Mathematical Society, 3:604–613, 1959.
•  M. R. da Silva, H. Ma, and A.-P. Zeng. Centrality, network capacity, and modularity as parameters to analyze the core-periphery structure in metabolic networks. Proceedings of the IEEE, 96:1411–1420, 2008.
•  E. Estrada. Open problem. the discriminant power of the subgraph centrality and other centrality measures. Discrete Applied Mathematics, 2013. To appear.
•  E. Estrada and N. Hatano. Communicability in complex networks. Physical Review E, 77(3):036111, 2008.
•  E. Estrada, N. Hatano, and M. Benzi. The physics of communicability in complex networks. Physics Reports, 514(3):89–119, 2012.
•  E. Estrada and J. A. Rodríguez-Velázquez. Subgraph centrality in complex networks. Physical Review E, 71(056103), 2005.
•  Ernesto Estrada. The Structure of Complex Networks: Theory and Applications. Oxford University Press, 2011.
•  L. C. Freeman. Centrality in social networks conceptual clarification. Social Networks, 1(3):215–239, 1979.
•  S. Gago Álvarez, J. Hurajová, and T. Madaras. Betweenness-selfcentric graphs. 2012. Preprint, available at http://upcommons.upc.edu/e-prints/handle/2117/15768.
•  A. Georgakopoulos. On walk-regular graphs and graphs with symmetric hitting times. arXiv:1211.5689, 2012.
•  R. Ghosh and K. Lerman. Rethinking centrality: The role of dynamical processes in social network analysis. arXiv:1209.4616, 2012.
•  C. D. Godsil and B. D McKay. Feasibility conditions for the existence of walk-regular graphs. Linear Algebra and Its Applications, 30:51–61, 1980.
•  L. Katz. A new status index derived from sociometric index. Psychometrika, pages 39–43, 1953.
•  M. E. J. Newman. Networks: An Introduction. Oxford University Press, 2010.
•  A. C. Paauwe. Distance regular graphs: Simply explained. 2007. Preprint, available at http://buzzard.ups.edu/courses/2007spring/projects/paauwe-paper-revised.pdf.
•  E. Pegg Jr. and E. W. Weisstein. Tutte 12-cage. MathWorld–A Wolfram Web Resource. Available at http://mathworld.wolfram.com/Tutte12-Cage.html.
•  G. Sabidussi. The centrality index of a graph. Psychometrika, 31:581–603, 1966.
•  J. Tits. Sur la trialité et certains groupes qui s’en déduisent. Institut des Hautes Études Scientifiques Publications Mathématiques, 2:14–60, 1959.
•  E. W. Weisstein. Chang graphs. MathWorld–A Wolfram Web Resource. Available at http://mathworld.wolfram.com/ChangGraphs.html.
You are adding the first comment!
How to quickly get a good reply:
• Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
• Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
• Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters   