DOI QR코드

DOI QR Code

Collective Betweenness Centrality in Networks

  • Gombojav, Gantulga (Department of Information and Computer Sciences, National University of Mongolia) ;
  • Purevsuren, Dalaijargal (Department of Information and Computer Sciences, National University of Mongolia) ;
  • Sengee, Nyamlkhagva (Department of Information and Computer Sciences, National University of Mongolia)
  • Received : 2022.05.03
  • Accepted : 2022.06.08
  • Published : 2022.06.30

Abstract

The shortest path betweenness value of a node quantifies the amount of information passing through the node when all the pairs of nodes in the network exchange information in full capacity measured by the number of the shortest paths between the pairs assuming that the information travels in the shortest paths. It is calculated as the cumulative of the fractions of the number of shortest paths between the node pairs over how many of them actually pass through the node of interest. It's possible for a node to have zero or underrated betweenness value while sitting just next to the giant flow of information. These nodes may have a significant influence on the network when the normal flow of information is disrupted. We propose a betweenness centrality measure called collective betweenness that takes into account the surroundings of a node. We will compare our measure with other centrality metrics and show some applications of it.

Keywords

I. INTRODUCTION

A large number of centrality metrics for various usages have been created over the years by network researchers. These centralities contributed significantly to analyzing and studying the structural properties of networks and individual characteristics of the nodes in complex networks. Centralities are used in identifying leaderships and influencer nodes in social networks [1], discovering important functionalities and structures in the decomposition of virus and brain networks [2-3], studying the spread of infectious diseases or information [4-6], segregating networks into small components [7], and analyzing the performance of Ad hoc networks [8-9].

Recent studies on degree centrality [10-13] introduce new collective centralities that are defined from the degree centrality of nodes in the neighborhood. Local centrality [10] uses the number of the nearest neighbors and the next nearest neighbors of a node as the neighborhood information. The local centrality of a node is the sum of the neighborhood information of the neighbors of that node. Collective Influencer [11-12] introduces a new perspective of centrality that the centrality of a node is defined from the collective centrality of all nodes on the ball having a certain distance from the node. A study in [13] proposes a centrality for identifying influential nodes which uses the coefficients based addition of the degree centrality of a node and the cumulative degree centrality values of the nearest neighbors of that node. A centrality score called Ing is proposed in [14] which assigns an initial value (i.e. degree centrality) to each node and repeatedly transforms the values of each node based on a transformation matrix. This process is repeated n times (algorithm parameter) to find the final Ing score.

Betweenness centrality (BC), which computes a rank for each node based on the role in communication between other nodes, is a popular measure to analyze networks [15- 18]. Betweenness centrality highlights bridging nodes that have a high impact on information exchange. The centralities based on bridging nodes are often neglected in previous studies [18]. In this study, we propose a collective betweenness centrality (collective BC), which computes the node centrality in the context of the collective bridging centrality. The collective betweenness value of each node is defined by the cumulative betweenness values of the nodes that are located within the ball, where the ball is the set of nodes within D distance (a given parameter) from a node in the network.

The experimental results show that the nodes identified by the collective BC can be influential spreaders under the SIR model. It is able to identify the nodes that become important nodes when the normal flow of information disrupts. We will study collective BC measurement and its correlation with another state-of-the-art centrality measurement in Section II. Section III covers an application of the collective BC with the susceptible-infected-recovered (SIR) model [4, 19] on five real-world networks. The final section gives a conclusion.

II. COLLECTIVE BETWEENNESS MEAS- UREMENT

The shortest path betweenness measures the amount of information that passes through the nodes of a given graph considering that the information travels in the shortest paths. Therefore, the betweenness value of a node is calculated as the fraction of the number of shortest paths passing through that node by the number of all the shortest paths. The betweenness value of 𝑣 is given by 𝑔(𝑣) function as shown in Equation 1 where 𝜎st(𝑣) denotes the number of shortest paths passing through node 𝑣 and 𝜎st denotes the number of shortest between 𝑠 and 𝑡 nodes. For simplicity, we may simply call shortest-path betweenness as betweenness (BC).

       (1)

In the shortest path betweenness, the nodes and edges involved in the current short path communication are highlighted and got high scores. There are some nodes and edges that are vital for communication if some disruption happens in the current short path communication.

Definition: some nodes and edges that are vital for communication if some disruption happens in the current short path communication are called “backbone nodes” for communication.

The proposed collective betweenness centrality ranks the backbone nodes in higher positions. The collective betweenness of a node is defined as the sum of betweenness values of the nodes located within 𝐷 distance away (measured in hops) from that node.

       (2)

In Equation 2, 𝑑(𝑢, 𝑣) denotes the distance between 𝑢 and 𝑣 nodes measured in hops. In other words, collective BC sums all the betweenness values within the ball centered at the node 𝑣 in addition to the betweenness of the center (node 𝑣). It gives more importance to the surroundings of the node since the betweenness sum of surrounding nodes usually exceeds the betweenness value of the node.

We used Spearman's 𝜌 correlation to analyze the relation of collective betweenness value with respect to three other centrality measurements -- degree centrality, closeness centrality, and betweenness centrality. For this measurement, the collective betweenness algorithm is run with D=1and D=2parameters on the datasets, and the correlation score are given for each centrality compared against collective betweenness for each dataset in a row in Table 1. Also, the mean correlation is calculated for each centrality to see the overall relationship of collective betweenness with others. The analysis shows that degree centrality and betweenness centrality are quite different from collective betweenness centrality for both D=1 and D=2 parameters. When D=1, the mean of coefficient 𝜌 is 0.62 and 0.57 for degree and betweenness centralities respectively. When D=2, the mean of coefficient 𝜌 is 0.63 and 0.58 for degree and betweenness centralities respectively. It seems that the distance parameter doesn’t affect the correlation between the centralities as the correlation increases on some networks and reduces on the other networks when the distance in- creases. Strangely enough, closeness centrality shows a high correlation to the collective betweenness scoring 0.87 and 0.89 for parameters D=1 and D=2 respectively. When D=1, the lowest coefficient for closeness centrality is 0.77 for netscience and the highest is 0.97 for email network. When D=2, the lowest is 0.70 for facebook and the highest is 0.98 for email network. This means that collective BC is quite correlated with closeness centrality for the most cases, although there may be some networks that make collective BC quite different from closeness centrality at least for these two depths.

Table 1. Correlation of collective BC against closeness (C), degree (D), betweenness (B) centrality.

Scatter plots of collective BC of the nodes in email and netscience networks are shown in Fig. 1. The collective BC (D=1) of the nodes are compared against their closeness, degree, and betweenness centralities. A median line is drawn for better clarification. The correlation coefficient in netscience network is 0.77, 0.46, and 0.47 for closeness, de- gree, and betweenness centralities respectively and it can be observed in Fig. 1. Most of the nodes in the third column (betweenness) are packed together around 0 betweenness showing that it’s not correlated with collective betweenness. The correlation with closeness can be seen in the first plot of netscience as we can see that when the closeness centrality increases, their collective BC also increases. High correlation with closeness, degree and betweenness centralities observed in email network of the figure matches with the data in Table 1 as their correlation coefficients are 0.97, 0.89 and 0.84. The highest correlated centrality is closeness and its plot in the first column of email network in the figure shows a nice concave shape as the closeness centrality value increases. The degree centrality of email network shows a linear dependency on the collective betweenness except little bit clustering along the vertical axis that means degree centrality doesn’t have clear ties with collective BC for those nodes. Obviously, the nodes with the same degree don’t necessarily have the same collective BC values. But, it’s still worthwhile to see that the correlation coefficient in this comparison is pretty much dependent on the structure of the network.

Fig. 1. Linear correlation between collective BC and other centralities. The horizontal axis is closeness centrality, degree centrality, or betweenness centrality. The vertical axis is collective BC (D=1).

III. EXAMPLES AND APPLICATIONS

3.1. Dataset

facebook is a Facebook community network where edges represent friendship and vertices represent one's profile. It was first used in [20]; grqc graph is a citation network of physicists working in General Relativity and Quantum Cos- mology. They can be downloaded from snap.stanford.edu; blog network contain communication networks of the blog owners on the Windows Live Spaces; email is a network made out of email communications of University Rovira i Virgili [21]; netscience is a co-authorship of scientists on network science. We used only the greatest connected component of this graph.

Table 2. Some characteristics (number of vertices (|𝑉|), number of edges (|𝐸| ), average degree (degree_avg), maximum degree (max_degree)) of networks.

3.2. Examples

In this subsection, we examine the difference between the shortest-path betweenness and collective betweenness centralities on example networks. We focus on the nodes getting a lower rank (score) in BC, but a higher rank in collective BC with D=1. Let’s see two simple examples which are presented in Fig. 2. BC and collective BC scores are computed and reported in Table 3

Fig. 2. Example networks for revealing the difference between collective BC and short-path BC.

Table 3. BC and collective BC scores of some nodes in two networks sketched in Fig. 2.

Node C in Fig. 2(a) has a zero BC score. In the real-world scenario, node C has somewhat participated in communication between nodes in Fig. 2(a). At least, if the link between node A and B is removed or de-activated, node C becomes vital for communication in this network. Although node E in Fig. 2(b) has four neighbors with high BC scores and is located in a more central position, it gets a lower BC score than its neighbors. Node E in Fig. 2(b) ranks first by the collective BC score. This node has a backbone role in communication. In addition, if some links in the short paths are removed or de-activated, e.g. link (A, B), node E gets a higher BC score. Our proposed collective BC highlights these backbone nodes.

A co-authorship network with 379 nodes is illustrated. We focus on the high-rank difference in two ranking methods, i.e., nodes getting lower rank (score) in BC but higher rank in collective BC with D = 1. The co-authorship network is presented in Fig. 3, where the nodes getting more than 120 ranks higher in collective BC are colored in red. The sizes of nodes represent the collective BC scores.

Fig. 3. The largest component of the co-authorship network. The nodes getting 120 (𝒏⁄𝟑) ranks higher in the collective BC than in BC are colored in red, where 𝒏 is the number of nodes. The sizes of nodes depend on their collective BC (D=1) values.

3.3. An Application for Identifying Influential Spreaders

The backbone nodes identified by our collective BC may serve as influential spreaders in networks. Susceptible-Infected-Recovered (SIR) model is one of the basic models that evaluate the spreading process in networks [4,19]. There are three compartments in this model: (i) Susceptible 𝑆(𝒕) represents the number of the people who are not infected by an infectious disease at the time 𝑡; (ii) Infected 𝐼(𝒕)represents the number of infected people (not yet recovered) at the time 𝑡 ; (iii) Recovered 𝑅(𝒕) represents the number of recovered people (they don't spread disease any more) from the disease at the time𝑡. At first, there is only one infected node. In each time step, each infected node is able to infect one of its susceptible neighbors with the probability 𝜆. We used 𝜆 = 1 for simplicity in evaluating the model just as described in [10]. An infected node is able to recover with the probability 1/degree_avg, where dgree_avg is the average degree of the graph shown in Table 1, at each step. When the number of infected nodes becomes zero, the process terminates. In order to investigate the influence of a single node that started the infection, the function 𝐹(𝑡) = 𝐼(𝑡) + 𝑅(𝑡) is used. 𝐹(𝑡) becomes constant after some time 𝑡and this value at 𝑡c becomes the influence of the starting node. The higher the value of 𝐹(𝑡c) the more influence (more devastation) the node has on the spread of disease or information in networks [10].

We examine the relation between the rank obtained by the collective BC and the influence score (𝑭(𝑡 = 10)) computed by the SIR model. The relation of the two measurements is compared with other centrality measures (closeness, degree, betweenness) on three networks (grqc, facebook, blogs). The results are presented in Fig. 4. The plot of degree centrality (fourth column) has long vertical lines that indicate nodes with the same degree can be a totally different influence in spreading, 𝐹(𝑡). The plot of betweenness centrality (fifth column) covers the almost whole area of the square. This indicates no clear relation between BC and influence in spreading, 𝐹(𝑡). Our collective BC (second and third column) is relatively better correlated with the node’s influence in spreading than that of degree and BC. In another words, the node with a high collective BC rank also has a tendency to have a high influence in spreading. The rank of collective BC (𝐷 = 2) is comparable with the rank of closeness centrality.

Fig. 4. The relation between the spreading influence of a node 𝑭(𝒕 = 𝟏𝟎) and the centrality score (closenness, collective BC with 𝑫 = 𝟏 and 𝑫=𝟐, degree, and betweenness) of the node.

IV. CONCLUSION

We have proposed a new centrality that is defined from the collectivity of betweenness centralities in the neighborhood. The proposed centrality highlights the nodes that are involved in the backbone role in the communication of networks. An application in identifying influential spreaders is examined under the Susceptible-Infected-Recovered (SIR) model. Experimental results show that the new centrality successfully identifies the influential spreaders in networks.

ACKNOWLEDGMENT

The research has received funding from the National University of Mongolia under grant agreement P2020-3978.

References

  1. L. Jain, R. Katarya, and S. Sachdeva, "Opinion leader detection using whale optimization algorithm in online social network," Expert Systems with Applications, vol. 142, p. 113016, Mar. 2020. https://doi.org/10.1016/j.eswa.2019.113016
  2. O. Sporns, C. J. Honey, and R. Kotter, "Identification and classification of hubs in brain networks," PLOS ONE, vol. 2, no. 10, p. e1049, Oct. 2007. https://doi.org/10.1371/journal.pone.0001049
  3. J. Santonja, K. Martinez, F. J. Roman, S. Escorial, M. Quiroga, and J. Alvarez-Linera, et al., "Brain resilience across the general cognitive ability distribution: Evidence from structural connectivity," Brain Struct Funct, vol. 226, no. 3, pp. 845-859, Apr. 2021. https://doi.org/10.1007/s00429-020-02213-4
  4. H. W. Hethcote, "The mathematics of infectious diseases," SIAM Rev., vol. 42, no. 4, pp. 599-653, Dec. 2000. https://doi.org/10.1137/S0036144500371907
  5. I. Cooper, A. Mondal, and C. G. Antonopoulos, "A SIR model assumption for the spread of COVID-19 in different communities," Chaos, Solitons & Fractals, vol. 139, p. 110057, Oct. 2020. https://doi.org/10.1016/j.chaos.2020.110057
  6. Q. Zeng, Y. Liu, M. Tang, and J. Gong, "Identifying super-spreaders in information-epidemic coevolving dynamics on multiplex networks," Knowledge-Based Systems, vol. 229, p. 107365, Oct. 2021. https://doi.org/10.1016/j.knosys.2021.107365
  7. S. Wandelt, X. Sun, D. Feng, M. Zanin, and S. Havlin, "A comparative analysis of approaches to network-dismantling," Scientific Report, vol. 8, no. 1, Sep. 2018.
  8. A. A. Aouiz, S. B. Hacene, and P. Lorenz, "Channel busyness based multipath load balancing routing protocol for ad hoc networks," IEEE Network, vol. 33, no. 5, pp. 118-125, Sep. 2019. https://doi.org/10.1109/mnet.2019.1900036
  9. J. H. Huh, Y. Kim, and K. Seo, "Power aware routing protocol in multimedia ad-hoc network considering hop lifetime of node," Journal of Multimedia Information System, vol. 1, no. 2, pp. 101-110, 2014.
  10. D. Chen, L. Lu, M. S. Shang, Y. C. Zhang, and T. Zhou, "Identifying influential nodes in complex networks," Physica A: Statistical Mechanics and Its Applications, vol. 391, no. 4, pp. 1777-1787, Feb. 2012. https://doi.org/10.1016/j.physa.2011.09.017
  11. F. Morone and H. A. Makse, "Influence maximization in complex networks through optimal percolation," Nature, vol. 524, no. 7563, Art. Aug. 2015.
  12. F. Zhu, "Improved collective influence of finding most influential nodes based on disjoint-set reinsertion," Scientific Report, vol. 8, no. 1, Sep. 2018.
  13. Y. Yang, X. Wang, Y. Chen, M. Hu, and C. Ruan, "A novel centrality of influential nodes identification in complex networks," IEEE Access, vol. 8, pp. 58742-58751, 2020. https://doi.org/10.1109/access.2020.2983053
  14. S. Xu, P. Wang, and J. Lu, "Iterative neighbour-information gathering for ranking nodes in complex networks," Scientific Report, vol. 7, no. 1, p. 41321, Feb. 2017. https://doi.org/10.1038/srep41321
  15. U. Brandes, "On variants of shortest-path betweenness centrality and their generic computation," Social Networks, vol. 30, no. 2, pp. 136-145, May 2008. https://doi.org/10.1016/j.socnet.2007.11.001
  16. C. A. Reis Pinheiro, M. Galati, N. Summerville, and M. Lambrecht, "Using network analysis and machine learning to identify virus spread trends in COVID-19," Big Data Research, vol. 25, p. 100242, Jul. 2021. https://doi.org/10.1016/j.bdr.2021.100242
  17. M. Ianni, E. Masciari, and G. Sperli, "A survey of big data dimensions vs social networks analysis," Journal of Intelligent Information System, vol. 57, no. 1, pp. 73-100, Aug. 2021. https://doi.org/10.1007/s10844-020-00629-2
  18. W. Liu, M. Pellegrini, and A. Wu, "Identification of bridging centrality in complex networks," IEEE Access, vol. 7, pp. 93123-93130, 2019. https://doi.org/10.1109/access.2019.2928058
  19. C. Castellano, S. Fortunato, and V. Loreto, "Statistical physics of social dynamics," Review of Modern Physics, vol. 81, no. 2, pp. 591-646, May 2009. https://doi.org/10.1103/RevModPhys.81.591
  20. J. Leskovec and J. Mcauley, "Learning to discover social circles in ego networks," Advances in Neural Information Processing Systems, 2012, vol. 25.
  21. R. Guimera, L. Danon, A. Diaz-Guilera, F. Giralt, and A. Arenas, "Self-similar community structure in a network of human interactions," Physcis Revew E, vol. 68, no. 6, p. 065103, Dec. 2003. https://doi.org/10.1103/PhysRevE.68.065103