This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

Digital Object Identifier

Generalization of Relative Change in a

Centrality Measure to Identify Vital
Nodes in Complex Networks
KODURU HAJARATHAIAH1 , MURALI KRISHNA ENDURI1 ,(Member, IEEE),
SATEESHKRISHNA DHULI2 , SATISH ANAMALAMUDI1 , AND LINGA REDDY
CENKERAMADDI3 ,(Senior Member, IEEE)
1
Department of Computer Science and Engineering, SRM University-AP, India.
2
Department of Electronics and Communications Engineering, SRM University-AP, India.
3
Department of Information and Communication Technology, University of Agder (UiA), Grimstad, 4879, Norway
Corresponding author: Linga Reddy Cenkeramaddi (email:[email protected])
This work was supported in part by the Indo-Norwegian Collaboration in Autonomous Cyber-Physical Systems
(INCAPS) of the INTPART International Partnerships for Excellent Education, Research and Innovation Program
from the Research Council of Norway under Project 287918.

ABSTRACT Identifying vital nodes is important in disease research, spreading rumors, viral
marketing, and drug development. The vital nodes in any network are used to spread information
as widely as possible. Centrality measures such as Degree centrality (D), Betweenness centrality
(B), Closeness centrality (C), Katz (K), Cluster coefficient (CC), PR (PageRank), LGC (Local
and Global Centrality), ISC (Isolating Centrality) centrality measures can be used to effectively
quantify vital nodes. The majority of these centrality measures are defined in the literature
and are based on a network’s local and/or global structure. However, these measures are time-
consuming and inefficient for large-scale networks. Also, these measures cannot study the effect
of removal of vital nodes in resource-constrained networks. To address these concerns, we propose
the six new centrality measures namely GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK.
We develop these measures based on the relative change of the clustering coefficient, degree, and
Katz centralities after the removal of a vertex. Next, we compare the proposed centrality measures
with D, B, C, CC, K, PR, LGC, and ISC to demonstrate their efficiency and time complexity. We
utilize the SIR (Susceptible-Infected-Recovered) and IC (Independent Cascade) models to study the
maximum information spread of proposed measures over conventional ones. We perform extensive
simulations on large-scale real-world data sets and prove that local centrality measures perform
better in some networks than global measures in terms of time complexity and information spread.
Further, we also observe the number of cliques drastically improves the efficiency of global centrality
measures.

INDEX TERMS Complex networks, influential nodes, local centrality, relative change in centrality.

I. INTRODUCTION to measure the spread of the disease. Authors proposed

The centrality measure effectively quantifies the net- a spreading scheme for viral marketing in [4] by using
work’s influential nodes. Identifying powerful or vital different centrality measures and some of the research
nodes is critical in many networking applications, in- insights presented to design marketing schemes. Various
cluding disease research [1], [2], fake news spreading [3], centrality measures have been proposed in [6] to identify
viral marketing [4], drug development, opinion monitor- the central nodes in large networks. It was also shown
ing [5], bio-logical field [6], fraud detection [7], structural that data can be reduced using machine learning meth-
design [8], social physics [9], and other fields [10]. In [1], ods to choose appropriate centrality measures. This is
authors analyzed the centrality measures for a disease the most meticulously researched fundamental concept
transmission network and shown that relatively simple in the network science. In the literature, the centrality
ego-network-based network measures can be adequate measure is defined in a variety of ways. According to [11],

VOLUME x, 2022 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

it is defined as the information or traffic contained in all the most significant invaders. This work is motivated
possible paths between pairs of nodes. These measures by these applications in the areas of cascading failures,
can be viewed as mathematical heuristics for identifying natural hazards, and network congestion.
prominent nodes in networks based on local or global
structural properties [12]. The centrality measures make II. RELATED WORK
certain assumptions about the process in which infor- This section briefly discusses the recent works on cen-
mation flows through a network [13]. In the last several trality measures in the literature. Authors have proposed
decades, many centrality methods have been proposed trust PageRank (TPR) [25], nearest neighborhood trust-
to find the seed nodes (vital nodes). Various centrality PageRank (NTPR) [26], extended cluster coefficient
measures, such as Degree centrality (D), Betweenness ranking measure (ECRM) [27], and normalised local
centrality (B) [14], Closeness centrality (C) [13], Katz centrality measure (NLC) [28] using local and global
(K) [15], PageRank (PR) [16], have been proposed in network information. The ECRM of a node is defined
the literature [17]. The centrality measures were broadly based on the correlation between a node and its neigh-
classified into three types: local measures, global mea- bors. By utilising the structure of the local network
sures, and random-walk measures [18]. The degree is around a node and the influence feedback from the
a local metric with low accuracy because it focuses on node’s closest neighbors, Zhao et al. [28] concentrated
first-order neighbors. The information from the global on normalised local centrality. Xin et al. [29] proposed a
network is ignored by the local metrics. The global heterogeneity-oriented immunization measure based on
metrics are betweenness and closeness, and these are individual heterogeneity and network topology factors.
considered the nodes’ global information. These global De et al. [30] designed an algorithm for detecting influen-
metrics, however, necessitate more computational time tial nodes in network dynamic systems using time series.
for large-scale networks [19]. A community-based mediator (CbM) [31] is presented as
PageRank is a random-walk measure that provides a metric for identifying influential nodes in a vast and
better performance and is appropriate for directed net- complex network, taking into account the entropy of a
works. All of these metrics rank nodes based on their random walk from a node to every community. The local
prominence in the network. A node with high central- and global centrality is proposed by the authors [32]
ity in a social network, for example, may represent a which is defined based on degree and shortest distance
powerful personality. The majority of these centrality between a pair of nodes. A centrality based on isolation
measures are defined by the number of paths that of vertex proposed by authors [33].
connect pairs of nodes, the shortest paths, betweenness, The degree cluster coefficient method (DCC) [34] is
degree, page rank, and so on. Several centrality measure- used for identifying influential nodes that takes into
ments are computed using local and global information account degree, clustering coefficient, and neighbors.
from network nodes. However, for large-scale complex The basic centralities and machine learning techniques
networks, these measures are time-consuming, costly, are used to find the vital nodes using SIR and the inde-
and inefficient. One of the interesting research direc- pendent cascade model [35]. The local relative change
tions in this area is developing centrality measures that of average shortest path (LRASP) [36] is proposed
consider node removal and neighborhood-level scenarios. depending on the network’s local structure. Based on
When influential nodes fail for any reason, nodes should the relative change in the average shortest path (ASP)
consider the alternate path or nearest neighbors, to find in the local network when the node is removed, the
the important nodes. These centrality measures find nu- LRASP measure for a node is calculated. From the
merous applications in many real-time networks. In [20], LRASP measure, we generalised the relative change in
authors proposed the centrality measures to study the centralities based on the global and local structures
National Airspace System, and airport network sub- to find the vital nodes. A new parallel algorithm is
jected to natural hazards. They have noted that node proposed in [23] to find all central nodes by obtaining
removal according to dynamic centrality measures can Breadth First Search trees. Authors proposed a belief
have faster collapse rates. Authors proposed a link cen- propagation and node reinsertion method in [37] to iden-
trality measure in [21] based on topological and electrical tify the vital nodes. A novel centrality measure based on
properties of power grid networks. They study the attack fuzzy concept is proposed in [38] which considered the
vulnerability of power grids in network failures. In [22], idea of inner structure of node’s box. In [39], authors
authors used centrality measures to reduce interference presented a comparative study of two vertex deleted
rate and congestion around the influential nodes in centrality measures namely Laplacian centrality and
Software-Defined Aerial Networks. Detecting the central algebraic centrality and proved that algebraic centrality
nodes which can be a base station can reduce the overall is easier to compute than Laplacian centrality. None of
energy consumption in multi-hop wireless networks [23]. these works focused on developing centrality measures
In [24], authors discussed the statistical measurements in which considers local and global information for node
constantly changing cooperative communities to identify removal cases using clustering coefficient, degree, katz
2 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

centralities. passes via vertex w is the shortest. The number of closed

In this work, we focus on deriving centrality measures triplets over the total number of triplets is used to cal-
to identify the vital node with maximal information culate the clustering coefficient centrality (CC) [42] [43].
spread and minimal time complexity. For this purpose, 2Nv
we propose six new centrality measures in this paper: CC(v) =
dv (dv − 1)
GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK.
These proposed methods focus on both local and global where Nv is number of links between neighbors of v, dv is
average structural information. We propose generalised degree of node v. In the Katz centrality (K) of a network,
centrality measures based on the relative change in the relative influence of each node is calculated by taking
degree, Katz, and clustering coefficient after node re- first level neighboring and next level neighboring nodes
moval. GRAD, GRACC, and GRAK are generalized that are connected with first level neighboring nodes.
global centrality measures based on the relative change The K of a node vi is computed as
in degree, clustering coefficient, and Katz, respectively. n
X
Similarly, local centrality measures are denoted as K(vi ) = α Ai,j K(vj )
LRACC, LRAD, and LRAK. The proposed methods j=1
were compared to the standard measures available in where α is dumping factor (considered to be less than
the literature, such as D, B, C, CC, K, PR, LGC, and largest eigen value).
ISC. PageRank (PR): A popular variation of the eigenvector
centrality technique, called PageRank centrality is pre-
A. ORGANISATION
sented in [44], [45]. In the Google search engine and
This paper is organized as follows. We define the basic
other commercial applications, PR is used to rank the
centrality measures in section III. Section IV presents
websites.
the proposed centrality measures and corresponding
1−α X P Rvt−1
algorithms. Section V is devoted to the data sets, P Rvt = +α i

spreading models, and assessment approaches utilized to n k vi

vi ∈Nv
prove the efficiency of our measures. Section VI includes
discussions and the experimental findings. Finally, Sec- where n represents vertices, Nv represents vertex v
tion VIII provides important insights, conclusions, and neighbors, α is jump probability, kvi is the number of
future directions. vertices to which the vertex vi points, and t is the
iterative parameter. The first term in PageRank is for
III. PRELIMINARIES regularising the PageRank. Here, the sum will converge
First, we define some benchmark centrality measures in to one when the second term reaches its maximum.
this section. Any graph or network is denoted by G, Local and Global Centrality (LGC): The LGC [32] is
formulated as G = (V, E), where V represents nodes defined based on the degree and shortest path between
and E represents edges. For computing the vital nodes a pair of vertices. It is a combination of the local and
in networks several centralities are represented in the global influence of a vertex. The LGC of node v is defined
existing work, such as degree (D), betweenness (B), as follows:
√
closeness (C), and clustering coefficient (CC) centralities dv X du + α
are defined as follows: The degree centrality (D) [40] is LGC(v) = ×
n d(u, v)
v̸=u
considered as number of direct ties between one vertex to
other vertices. The average length of the shortest path where dv represents the degree of node v, n is total
connecting a node to all other nodes in the graph is number of nodes, d(u, v) is the shortest distance between
known as Closeness centrality (C) [41]. The C of node v u and v, and α is the parameter and range between 0
can be defined as and 1. First part indicates the local influence and second
1 part indicates the global influence.
C(v) = P Isolating Centrality (ISC): The ISC [33] of a node
d(v, u)
u∈V is defined as the product of its degree and isolated
where d(v, u) indicates the shortest path distance be- coefficient. The ISC of a node is calculated as:
tween u, v nodes. The betweenness centrality (B) [14]
ISC(v) =| Nv ∩ Dδ | ×dv
is global centrality and it considers the shortest path
through the node. The B of node w can be defined as where Nv is the neighbors of v and Dδ is set of nodes
X duv (w) with degree δ, and dv is degree of node v. The average
B(w) = shortest path (ASP) [43] is defined as the shortest paths
duv
u̸=v̸=w∈V for all potential network pairs along with the average
where duv is the distance from the vertex u to vertex number of steps. The ASP measures how well informa-
v, duv (w) is the path between vertices u and v that tion is effectively transmitted from powerful nodes to all
VOLUME x, 2022 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

other nodes in the network. The ASP is computed for a summary of our proposed methods and existing methods
graph G, as P are shown in Table 1.
duv Consider any centrality C , for every vertex we can
u̸=v∈V
ASP [G] = find the values by using this centrality C and also we
N (N − 1) can find the average centrality value for graph and it is
where duv denotes the shortest path from nodes u and denoted as AvgC . We can define two measures, global
v. The relative change in ASP (RASP) [46] can be and local, based on the effect of relative change. It means
computed as we investigate the effect of relative change of AvgC once
the vertex is removed. The global measure is defined as
|ASP [G′k ] − ASP [G]|
RASP [k] = , k = 1, 2, · · · , N follows:
ASP [G]
|AvgC [G′v ] − AvgC [G]|
In a graph G, v is a vertex, consider a neighbourhood GRAvgC (v) = (1)
AvgC [G]
level L. The LRASP [36] is a local centrality and
calculated nodes influence based on the network’s local where G′v is a network after removing the vertex v
structure. This LRASP metric for a node is based on from network G. The centrality GRAvgC is global for
the relative change in the average shortest path in the finding the centrality of a vertex in the entire network.
local network when the node is eliminated. NL (v) is Process of finding the GRAvgC of vertex v of graph G
the set of all neighbors up to the level L with vertex given in Algorithm 1. The global measure defined above
v. RASP of GNL (v) will be calculated for an induced
subgraph GNL (v) of graph G with vertex set NL (v). Algorithm 1: Algorithm for finding global mea-
The summary of the methods is shown in Table 1. The sure GRAvgC (v) for a vertex in graph G
LRASP is computed as follows: Input: Graph G = (V, E), vertex v
|ASP [GNL (v) \ v] − ASP [GNL (v) ]| Output: Global measure (GRAvgC (v)) of a
LRASPL [v] = , vertex v
ASP [GNL (v) ]
-21 begin
-22 V = nodelist, E = edgelist
A. CONTRIBUTIONS:
-23 for every vertex v in V do
1) Firstly, we propose a generalised centrality measure
-24 find C (v)
by using the relative change in any centrality. For
that, the effect of the centrality measure after -25 find AvgC (G)
deleting a vertex has been exploited. -26 similarly find AvgC [G′v ] /* G′v is a graph
2) Secondly, we propose six centralities such as after removing a vertex v from G */
|AvgC [G′v ]−AvgC [G]|
GRACC, LRACC, GRAD, LRAD, GRAK, and -27 find GRAvgC (v) = AvgC [G]
LRAK and compared with the conventional mea- -28 return (GRAvgC (v))
sures D, B, C, CC, K, PR, LGC, and ISC. Proposed
measures are extremely useful to study the complex
networks with less computational complexity. based on the global structure and for this measure we
3) Finally, to verify the maximum information spread, need entire network information. Now we define a local
we test our centrality measures on real-world data measure based on local structure of graph and for this
sets using SIR and IC models. measure we need local network information. Consider
Next we define generalization of the centrality measure the neighbourhood level L for a vertex v in G and NL (v)
using the relative change in the centrality. is the vertex v of the neighbors up to the level L in
the graph. The level L can range from 0 to the graph’s
IV. GENERALIZATION OF RELATIVE A CHANGE IN diameter. Let us define the local measure which is as
CENTRALITY follows:
In this section, we generalize the centrality measure |AvgC [GNL (v) \ v] − AvgC [GNL (v) ]|
LRAvgCL (v) = ,
based on the relative change in the centrality. We AvgC [GNL (v) ]
propose this generalization of the centrality measure of a (2)
vertex in the network by using the effect of the centrality where GNL (v) \v is a graph GNL (v) after deleting a vertex
measure after removing a vertex. The centralities of a v. Assume a neighbourhood level L for a vertex v in G,
node defined in the literature are based on network local NL (v) is the vertex v of the neighbors up to the level
structure and global structure. But we defined the local L in the graph. Find the AvgC of induced subgraph
and global centrality measures of a node by observing GNL (v) of graph G with vertex set NL (v). The centrality
the effect of the relative change in any centrality once the LRAvgC is local for finding the centrality of a vertex
node is removed. Furthermore, we list local and global by involving only neighboring vertices up to the level
centrality measures based on this generalization. The L. Process of finding the LRAvgCL of a vertex v of
4 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

Algorithm 2: Algorithm for finding local measure where K(v) indicates the Katz centrality value of a node
LRAvgCL (v) for a vertex in graph G v. The global relative change in average Katz centrality
Input: Graph G = (V, E), vertex v and level L can be calculated as:
Output: Local measure (LRAvgCL (v)) of a |AK[G′k ] − AK[G]|
GRAK(v) = ,v ∈ V
vertex v AK[G]
-21 begin Local relative change in average Katz centrality of a
-22 V = nodelist, E = edgelist, n := len(V ); vertex v of G defined as follows:
-23 for i from 1 to L do
|AK[GNL (v) \ v] − AK[GNL (v) ]|
-24 find the neighbours of vertex v for each LRAKL (v) = ,
level i and add to the set NL (v) by using AK[GNL (v) ]
edgelist E.
C. CENTRALITY C IS DEGREE
-25 Consider H = GNL (v) /* find induced If we consider centrality C as degree (D) [48] the average
subgraph of G with set vertex set NL (v) by degree is calculated as:
using edge list E */ P
-26 find LRAvgCL (v) = |AvgC [H\v]−AvgC [H]| dv
AvgC [H] v∈V
-27 where H \ v is a graph obtained after AD[G] = (5)
N
removing a vertex v from V (H) and we can
where dv denotes the degree of a node v. The global
also find AvgC of a graph.
relative change in average degree of a node v in graph
-28 return (LRAvgCL (v))
G can be determined as:
|AD[G′v ] − AD[G]|
GRAD(v) = ,v ∈ V
AD[G]
graph G is given in Algorithm 2. Next we list local and Local relative change in average degree of a node v in
global centrality measures based on this generalization. graph G defined as follows:
We consider the centrality C as clustering coefficient,
|AD[GNL (v) \ v] − AD[GNL (v) ]|
degree and Katz centrality which are shown in Table 1. LRADL (v) =
AD[GNL (v) ]
A. CENTRALITY C IS CLUSTERING COEFFICIENT
If we consider centrality C as clustering coefficient
(CC) [47] then the average clustering coefficient (ACC)
is calculated for graph G as:
P
CC(v)
v∈V
ACC[G] = (3) FIGURE 1. A Toy Graph with 12 vertices and 14 edges.
N
where CC(v) indicates the clustering coefficient of a Example: We illustrate these proposed centralities by
node v. The global relative change in average clustering using toy example. Simple graph with 12 vertices and
coefficient of a vertex v is defined as: 14 edges is given in the Fig. 1. For the network given in
the Fig. 1, we find the centrality D, B, C, CC, K, PR,
|ACC[G′v ] − ACC[G]|
GRACC(v) = , LGC, ISC, GRACC, LRACC, GRAD, LRAD, GRAK,
ACC[G] and LRAK values for every vertex which are given in
Local relative change in average clustering coefficient for the Table 2. While finding the local measures (LRACC,
a vertex v of G defined as follows: LRAD and LRAK), we consider L is half of the diameter.

|ACC[GNL (v) \ v] − ACC[GNL (v) ]| D. COMPLEXITY OF FINDING GRAV GC AND

LRACCL (v) = ,
ACC[GNL (v) ] LRAV GC
Let us consider the network G = (V, E), where |V | = n
denotes the number of nodes and |E| = m represent
the number of edges. For a graph G, consider the
B. CENTRALITY C IS KATZ CENTRALITY
computational complexity for evaluating the relative
If we consider centrality C as Katz centrality (K) [15] change in AvgC for all vertices as O(f (n)). The com-
then average Katz centrality (AK) of graph is deter- plexity for finding the global measure GRAvgC in
mined as: P algorithm 1 is O(f (n)). The computational complexity
K(v) for local centrality LRAvgC mentioned in algorithm 2
v∈V
AK[G] = (4) is O(n2 + f (|NL (v)|)). Here, v denotes number of nodes
N
VOLUME x, 2022 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

TABLE 1. GRAvgC and LRAvgC are proposed centralities, where C stands for D (Degree), K (Katz), CC (Clustering Coefficient), and ASP (Average
Shortest Path). The Proposed centralities in this work are colored in red.

Centrality C AvgC GRAvgC (v) LRAvgCL (v)

P
dv
v∈V |AD[G′v ]−AD[G]| |AD[GN (v) \v]−AD[GN (v) ]|
L L
D AD[G] = GRAD(v)= LRADL (v) =
N AD[G] AD[GN (v) ]
P L
K(v)
v∈V |AK[G′ ]−AK[G]| |AK[GN (v) \v]−AK[GN (v) ]|
K AK[G] = GRAK(v)= k LRAKL (v) = L L
N AK[G] AK[GN (v) ]
P L
CC(v)
v∈V |ACC[G′v ]−ACC[G]| |ACC[GN (v) \v]−ACC[GN (v) ]|
L L
CC ACC[G] = GRACC(v)= LRACCL (v) =
PN ACC[G] ACC[GN (v) ]
L
duv
u̸=v∈V |ASP [G′v ]−ASP [G]| |ASP [GN (v) \v]−ASP [GN (v) ]|
L L
ASP ASP [G] = RASP (G) = [46] LRASP (G) = [36]
N (N −1) ASP [G] ASP [GN (v) ]
L

TABLE 2. D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local and Global Centrality), ISC (Isolating
Centrality), GRACC (Global Relative change in Average Clustering Coefficient), LRACC (Local Relative change in Average Clustering Coefficient), GRAD (Global
Relative change in Average Degree), LRAD (Local Relative change in Average Degree), GRAK (Global Relative change in Average Katz), and LRAK (Local
Relative change in Average Katz) centralities values at a vertex for a graph are shown in Figure 1. Top three influenced nodes are represented in red color.

Vertex(v) 1 2 3 4 5 6 7 8 9 10 11 12
D(v) 1 3 1 4 2 3 3 3 1 3 2 2
B(v) 0 0.345 0 0.491 0 0.545 0.55 0.564 0 0.327 0 0
C(v) 0.275 0.367 0.275 0.458 0.333 0.478 0.407 0.44 0.314 0.355 0.275 0.275
K(v) 0.285 0.291 0.285 0.293 0.288 0.291 0.291 0.291 0.285 0.291 0.288 0.288
CC(v) 0 0 0 0.333 1 0.333 0.667 0 0 0.333 1 1
P R(v) 0.0460 0.118 0.0460 0.129 0.067 0.097 0.096 0.106 0.043 0.105 0.073 0.073
LGC(v) 0.578 2.315 0.578 3.725 1.451 2.752 2.521 2.577 0.623 2.316 1.253 1.253
ISC(v) 0 6 0 0 0 0 0 3 0 0 0 0
GRACC(v) 0.091 0.169 0.091 0.454 0.065 0.091 0.454 0.403 0.091 0.454 0.454 0.454
LRACC(v) 0.010 0.558 0.010 0.091 0.247 0.091 0.091 0.169 0.273 0.636 0.455 0.455
GRAD(v) 0.013 0.143 0.013 0.221 0.065 0.143 0.143 0.143 0.013 0.143 0.065 0.065
LRAD(v) 0.034 0.111 0.034 0.174 0.030 0.143 0.091 0.167 0.051 0.111 0.091 0.091
GRAK(v) 0.045 0.042 0.045 0.046 0.046 0.047 0.045 0.043 0.045 0.042 0.043 0.043
LRAK(v) 0.202 0.148 0.202 0.043 0.145 0.047 0.044 0.048 0.149 0.151 0.264 0.264

TABLE 3. Basic properties of six data sets.

Network Nodes Edges Max Degree Avg. Degree Diameter Avg. Clustering Coeff.
USAir97 332 2126 139 12.8 6 0.625
bio-celegans 453 2025 237 8.94 7 0.646
ca-netscience 379 914 34 4.82 17 0.741
web-polblogs 643 2280 165 7.09 10 0.232
email-univ 1133 5451 71 9.62 8 0.220
Fb_Pages 14113 52310 215 7 9 0.239

and NL (v) represents neighbors of vertex v to the level the time complexity of the proposed centralities in the
L. Finding the induced subgraph in a given graph takes Table 4.
run time O(n2 ), and finding the LRAvgC value for this
induced subgraph costs f (|NL (v)|) where |NL (v)| is the TABLE 4. Time complexity of finding AvgC , GRAvgC and LRAvgC
where C is D, K, CC, and ASP, where dmax is the maximum degree of a
size of the induced subgraph. The level L can range from network, n (m) number of vertices (edges) of graph, and L represents
0 to the graph’s diameter. An alternative way, if L is the neighborhood level.
graph’s diameter, then LRAvgC equals GRAvgC . In
worst-case scenario, the |NL (v)| value is n. In this paper, C T (AvgC ) T (GRAvgC ) T (LRAvgCL )

we assume that level L is half the diameter of the graph.

D O(n2 ) O(n2 ) O(max{n2 , |NL (v)|2 })
In the given graph for calculating centrality, LRAvgC
for each vertex, takes O(n(max{n2 , f (|NL (v)|)})). This K O(mn) O(mn) O(max{n2 , |NL (v)|3 })
time complexity is defined by the local structure of
O(n2 d2 O(n2 d2 O(max{n2 , |NL (v)|4 })
each vertex as max{n2 , f (|NL (v)|)}. Computing global CC max ) max )

measure GRAvgC takes more runtime when compared ASP O(n3 ) O(n3 ) O(max{n2 , |NL (v)|3 })
to local LRAvgC measure. Computing the LRAvgC
measure for any graph is very efficient. We present
The neighborhood level L values can be varied from
6 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 2. Degree distribution for USAir97, bio-celegans, ca-netscience, web-polblogs, email-univ, and Fb_Pages data sets.

real value 1 to diameter. We have examined that L infected (I), and recovered (R). For the implementation
is equal to half of the diameter that covers the more process, the top−10 nodes are picked for infection nodes
local neighbors’ information with less time complexity. based on a centrality score. For every time instant, an
From our simulation experiments, we have also ob- infected node tries with a chance to infect one of its
served that this particular L value maintains a trade- neighbors with probability β. Simultaneously, there is
off between time complexity and maximum informa- a high chance of success for every infected node to get
tion spread. One more interesting finding is when L recovered with a probability γ, if successful, it will never
converges to diameter, it will be a global measure. In be infected again and will no longer infect additional
our simulations, we have observed that LRACC(v) = susceptible nodes. Further, the completion of the process
GRACC(v), LRAD(v) = GRAD(v), and LRAK(v) = indicates that the network has no infected nodes. The
GRAK(v) when L is the diameter of the network. diffusion process was replicated 100 times in this study.
The infection rate β is considered to be in the range of
V. IMPLEMENTATION 0.1 to 0.3 for the simulations in the SIR model.
This section illustrates six real distinct networks to Independent cascade model (IC) is an information
evaluate the performance of our methods. Later SIR, diffusion model [52], [53]. The IC model is a dynamical
the Independent cascade model, and Kendall rank cor- information propagation approach in which data travels
relation are described to analyze the results. through a cascade across the network. Considering the
average of a massive number of Monte Carlo simulations
A. DATA the expected spread of a given seed set is computed.
Six real networks are used for simulations such as US- Nodes can exist in either an active or passive. (i)
Air97, bio-celegans, ca-netscience, web-polblogs, email- Active signifies the already influenced node by the data
univ, and Fb_Pages. These real networks are down- available in diffusion. (ii) Inactive signifies that the node
loaded from [49]. The basic details of the data sets are is entirely ignorant of the information. The IC and SIR
summarised in the Table 3. The degree distribution plots models are used to examine our proposed measures.
for six datasets are shown in Fig. 2. The degree dis-
tribution (neighbor distribution) is the most significant C. KENDALL COEFFICIENT
characteristic of a network structure. It is a rank correlation metric that measures how com-
parable the data’s orderings are when ranked by each of
B. SPREADING MODELS the variables [54]–[56]. A similar rank in observations
The SIR (Susceptible-Infected-Recovered) model is one results in a high Kendall correlation among the two
of the epidemic model [50], [51]. To assess the techniques variables where as dissimilar rank in observations results
in this report, we apply the SIR model with minimal in low correlation between two variables. Consider a set
contact. Each node in the SIR model must be in one of observations (a1 , b1 ), · · · , (an , bn ) in which A, B are
of these three states. state of being susceptible (S), two random variables and (xi ), (yi ) are unique values.
VOLUME x, 2022 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

TABLE 5. Rank Correlation (τ ) values of various ( Where D (Degree), B (Betweenness), C (Closeness), CC (Clustering coefficient), K (Katz), PR (PageRank),
LGC (Local and Global Centrality), ISC (Isolating Centrality), GRACC (Global relative change in average clustering coefficient), LRACC (Local relative change in
average clustering coefficient), GRAD (Global relative change in average degree), LRAD (Global relative change in average degree), GRAK (Global relative change
in average Katz), and LRAK (Local relative change in Katz)) centralities.

Networks D B C CC K PR LGC ISC GRACC LRACC GRAD LRAD GRAK LRAK

USAir97 0.002 0.123 0.150 0.118 0.117 0.031 0.093 0.862 0.044 0.045 0.054 0.211 0.240 0.029
bio-celegans 0.070 0.058 0.014 0.003 0.020 0.029 0.018 0.982 0.030 0.116 0.033 0.111 0.007 0.050
ca-netscience 0.233 0.320 0.144 0.180 0.205 0.235 0.226 0.821 0.080 0.175 0.241 0.144 0.099 0.123
web-polblogs 0.154 0.125 0.166 0.151 0.030 0.039 0.052 0.715 0.077 0.108 0.177 0.148 0.060 0.026
email-univ 0.481 0.423 0.520 0.141 0.447 0.389 0.450 0.727 0.053 0.489 0.353 0.486 0.431 0.014
Fb_Pages 0.012 0.010 0.001 0.035 0.007 0.010 0.008 0.777 0.002 0.169 0.022 0.280 0.005 0.002

For any observations (ai , bi ) as well as (aj , bj ), in which the new rankings (new measures) with basic ranking
i < j are concordant if both (ai > aj ) and (bi > bj ) (X = 1, 2, 3, · · · , N ) [38], [57], [58]. Table 5 shows that
holds or (ai < aj ) and (bi < bj ) holds. Else they are correlation between new measures and basic centralities
discordant. The Kendall’s τ coefficient is computed as for all vertices. In Fig. 3, we show the correlation at
C −ND
τ = 1N(n(n−1)) , where NC denotes the total number of every value of x where x is from 1 to the number of
2
concordant pairs and ND denotes the total number of vertices of network. In the network, we compute rankings
discordant pairs respectively. The range of coefficients of every node by using centrality measures. In USAir97
must be within the acceptable range −1 ≤ τ ≤ 1, for data set, the correlation between LRACC and LRAD
the denominator to be in the total number of possible is almost close and LRACC is not correlated with other
combinations in the pair. The coefficient is said to have centralities. In bio-celegans, CC and LRAK, D and B are
a specific value of 1 if rankings there is a perfection very closely related respectively. In ca-netscience, B and
in the agreement among the two ranking whereas the GRAD are closely correlated. Our proposed centralities
coefficient value is −1 if there is a discrepancy between (GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK)
the two ranks which is perfect. We would anticipate the are not close to basic centralities in web-polblogs, shown
coefficient to be close to zero if R and S are distinct. in Fig. 3. The proposed centralities’ performance is
not close to the conventional measures in email-univ
VI. RESULTS AND DISCUSSION and Fb_Pages networks. Initially, GRACC and LRAD
We show the results from various datasets in this section. centralities are highly correlated with each other. Similar
Initially, we display the correlation between proposed results were observed for GRAK and LRAK centralities.
centrality and basic centralities. We explain that the As shown in Fig. 3, other centralities are not correlated
spread of information grows as the centrality of a in Fb_Pages network.
network node’s value rises. We examined cumulative
infected nodes for D, B, C, CC, K, PR, LGC, ISC, B. SPREADING ABILITY WITH CENTRALITY VALUE
GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK Considering the count of 100 simulations in the SIR
centralities using the SIR and independent cascade model, we study the relationship between the spread
models. For these centralities, we observe the pattern ability and the centrality value of the node in this
of maximal effect with varying infection rates. section. Various centrality methods are assessed for
the centrality value of every node. The node with the
A. RANK CORRELATION OF GRACC, LRACC, GRAD, highest centrality value is considered as an infected
LRAD, GRAK, AND LRAK WITH BASIC CENTRALITIES node. With 100 times SIR model simulations (described
We define the global and local measures such as in section V), the total number of infected people is
GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK calculated. The infection probability β is set within
in section IV. We investigate the proposed centralities the range 0.1 to 0.3. If the infection probability goes
close with any existing centralities by using the Kendall beyond 0.4, then most of the people in the network will
coefficient. On six real networks, we show the correla- be infected. Using centrality methods such as GRACC,
tion of GRACC, LRACC, GRAD, LRAD, GRAK, and LRACC, GRAD, LRAD, GRAK, LRAK, ISC, LGC,
LRAK with D, B, C, CC, K, PR, LGC, and ISC cen- PR, K, CC, C, B, and D, a comparison between the
tralities. Using the network’s centralities, we calculate centrality of node’s value and the infection rate is graph-
the ranking of each vertex. For each top N vertices, ically plotted. In the Fig. 4, and 5 it is observed that
we provide correlation graphs of the GRACC, LRACC, with an increase in infection rate, there is an increase
GRAD, LRAD, GRAK, LRAK with fundamental cen- in the centrality value. The experimental results, in
trality methods in Fig. 3, here N represents set of values Fig. 5(a), show that the proposed methods LRACC,
i.e. N = {1, 2, · · · , n} and n denotes total number LRAD, and LRAK spread more information compared
of nodes in a network. We find correlation between to other basic centrality methods. More information is
8 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 3. Rank Correlation with basic centralities for six data sets. Where D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K (Katz),
PR (PageRank), LGC (Local and Global Centrality), ISC (Isolating Centrality), GRACC (Global Relative change in Average Clustering Coefficient), LRACC (Local
Relative change in Average Clustering Coefficient), GRAD (Global Relative change in Average Degree), LRAD (Local Relative change in Average Degree), GRAK
(Global Relative change in Average Katz), and LRAK (Local Relative change in Average Katz).

being disseminated through GRACC and C centralities. centrality. In the web-polblogs data set, LRACC, LRAD,
As shown in Fig. 4(b), C, LRACC, and LRAK central- and LRAK spread more than other centralities, as shown
ities spread more information than D, B, CC, K, PR, in Fig. 5(d). Later, GRACC and C centralities transmit
LGC, ISC, GRACC, GRAD, LRAD, GRAK. Later more more information in the network. In Fig. 5(e), LRACC
information is being disseminated through GRAD and and LRAK centrality infection rates are higher. Later
GRAK centralities. As shown in Fig. 5(c), our proposed C and GRACC have information transmission that is
local centralities are transforming information more higher than other centralities. The spread of information
than other centralities. In this dataset, all centralities increases with the node’s influence. In the Fb_Pages
are transmitting more information except betweenness network, GRAD, LRAK, and C centralities have higher

VOLUME x, 2022 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 4. Using SIR model, Centrality value with infection rate for six networks in column wise. Where D (Degree), B (Betweenness), C (Closeness), CC
(Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local and Global Centrality), and ISC (Isolating Centrality).

infection rates than other centralities as shown in the and basic centrality methods (D, B, C, CC, K, PR),
Fig. 5(f ). the latest measures (LGC, ISC), the most influential
top − 10 nodes were calculated. The top − 10 vertices
C. CUMULATIVE INFECTED NODES FOR GRACC, which have high centrality are infected initially in the
LRACC, GRAD, LRAD, GRAK, AND LRAK WITH BASIC SIR model. In the next stage, the seed nodes of the
CENTRALITIES surrounding/neighboring vertices are infected with in-
This section displays the overall infected nodes, as fection probability β. The predicted value for infection
well as the effect of distributing the information af- probability lies within the range of 0.1 to 0.3. Most
ter being infected by the top − 10 influential (vital of the people in the network will be infected if the
nodes) nodes. Using the proposed centrality methods infection probability goes beyond 0.5. After a certain
(GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK) time period, anyone who becomes infected can be re-

10 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 5. Using SIR model, Centrality value with infection rate for six networks in column wise. Where GRACC (Global Relative change in Average Clustering
Coefficient), LRACC (Local Relative change in Average Clustering Coefficient), GRAD (Global Relative change in Average Degree), LRAD (Local Relative change
in Average Degree), GRAK (Global Relative change in Average Katz), and LRAK (Local Relative change in Average Katz).

covered at a specific rate γ, which is defined as 1. The spread information faster in Fb_Pages network as shown
100 simulations resulted in cumulative infected nodes in the Fig. 6. We have observed a similar phenomenon for
on an average. We consider the 10 time steps in Fig. 6 the independent cascade method. Using the independent
which depicts the results. In bio-celegans, ca-netscience, cascade model (IC model), we illustrate the average
web-polblogs, our proposed centralities LRAD, LRAK, number of nodes that are information retrieved with
and GRAD all have more spread ability than the other different time scales in Fig. 7. Different centrality mea-
centralities. In the USAir97 data set, the betweenness surements are used to construct the seed nodes, which
is performed well and also our centrality LRACC is are input to IC model. 1000 iterations were employed for
performed well. Initially, up to some intervals, the degree simulations in the IC model. In the USAir97 network,
is performed well. Later, our centrality measure LRACC the average information spread is more for our proposed
reaches the top along with the betweenness. In the methods such as LRACC, and LRAD compared to
bio-celegans data set, LRACC has more spread ability other centralities. In the bio-celegans network, C, CC,
than the other centralities. Later, betweenness centrality and GRAK spread more information than D, B, K,
also transmits more information, as shown in Fig. 6. GRACC, LRACC, GRAD, LRAD, and LRAK. The
GRAD is in top position for spreading information results are shown in Fig. 7. GRAD has a better average
compared to other centralities in web-polblogs. Later, information spread than the other methods in the ca-
Katz centrality is in the ability to spread more. Our netscience network. GRAD, GRAK, and GRACC have
centralities have the ability to spread more in web- good information spread compared to other centralities
polblogs, and some of the basic centralities also have in the web-polblogs, the results shown in Fig. 7. In email-
the ability to spread more. GRAD centrality spreading univ network, GRAK centrality average information
ability is greater when compared with other centralities spread is more. As shown in the Fig. 7, GRAK centrality
in the email-univ. GRAK, GRACC, and C centralities information is more effective over other measures in
VOLUME x, 2022 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 6. 100 simulations of the SIR model’s cumulatively infected nodes for the USAir97, bio-celegans, ca-netscience, web-polblogs, email-univ, and Fb_Pages
networks. The first 10 nodes listed are infected nodes, as determined by proposed centralities and different centralities.

Fb_Pages network for both SIR and IC models. of infection probability was estimated in the range of 0.1
and 0.5 by using the SIR model (100 simulations). If the
infection probability goes beyond 0.5, then most of the
D. NORMALIZED MAXIMUM INFLUENCE FOR
people in the network will be infected. When compared
GRACC, LRACC, GRAD, LRAD, GRAK, AND LRAK
with conventional centralities, it was observed that our
WITH BASIC CENTRALITIES
proposed centrality showed the highest infection pop-
This section displays the assessment of infection spread ulation at different levels of infection probability. In
ability for the top-10 most influential nodes with differ- this part, we use the IC model to discuss maximum
ent infection rate. These nodes are discovered by the information spread with various infection probabilities
D, B, C, CC, K, PR, LGC, ISC, GRACC, LRACC, (range from 0.1 to 0.3). The results in Fig. 8, show 1000
GRAD, LRAD, GRAK, and LRAK centrality methods. iterations of IC model simulations. In Fig. 8, we dis-
The important nodes have the potential to propagate de- played normalized infection (maximum) with different
pending on the information in the networks. Evaluation
12 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 7. Using the independent cascade model (IC model) with top − 10 seed nodes, we calculated the average information spread among GRACC, LRACC,
GRAD, LRAD, GRAK, and LRAK with other centralities for the six networks Where D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K
(Katz), PR (PageRank), LGC (Local and Global Centrality), ISC (Isolating Centrality), GRACC (Global Relative change in Average Clustering Coefficient), LRACC
(Local Relative change in Average Clustering Coefficient), GRAD (Global Relative change in Average Degree), LRAD (Local Relative change in Average Degree),
GRAK (Global Relative change in Average Katz), and LRAK (Local Relative change in Average Katz).

infection probabilities. On the USAir97 network, our in email-univ and Fb_Pages networks. In Fig. 9, using
methods GRAD, GRAK along with B performed well IC model, we describe maximum information spread
when evaluated with other centralities. GRAK, GRAD, with varied probability infection. In the USAir97 net-
GRAK, LRAK have highest spread ability than rest of work, D and GRACC spread maximum information
the centralities on the bio-celegans. GRAD, K are posi- than B, C, CC, K, LRACC, GRAD, LRAD, GRAK,
tioned top on the ca-netscience data set when compared and LRAK. In the bio-celegans, the LRACC, LRAK,
with rest of the centralities. Similarly GRACC, LRAK and D centrality methods transfer more information. In
are top positioned on the web-polblogs, which are shown the ca-netscience data set, GRAD showed the highest
in the Fig. 8. As shown in the Fig. 8, faster information information spread, which is shown in the Fig. 9. In web-
spreading is observed for GRAK and GRAD centralities polblogs, our centrality methods are at the top, which

VOLUME x, 2022 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 8. Using the SIR simulations, the maximum information spread of the top − 10 influenced nodes of networks with varied infection probabilities was
normalized. Where D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local and Global Centrality), ISC
(Isolating Centrality), GRACC (Global Relative change in Average Clustering Coefficient), LRACC (Local Relative change in Average Clustering Coefficient),
GRAD (Global Relative change in Average Degree), LRAD (Local Relative change in Average Degree), GRAK (Global Relative change in Average Katz), and
LRAK (Local Relative change in Average Katz).

indicates that information spread is greater. GRAK and LRACC, GRAD, LRAD, GRAK, and LRAK on US-
GRAD centralities exhibit faster information spread for Air97, bio-celegans, ca-netscience, web-polblogs, email-
email-univ and Fb_Pages networks, whereas GRACC univ, and Fb_Pages datasets. We have observed that the
centrality shows better performance in Fb_Pages as performance of GRACC, GRAK, and GRAD measures
shown in Fig. 9. Similar observations are clearly seen is good for ca-netscience, email-univ networks, web-
from both the methods of SIR and IC. polbogs, and Fb_Pages. Similarly, the performance of
LRACC, LRAD, and LRAK is better for USAir97 and
VII. COMPARISON OF PROPOSED CENTRALITIES bio-celegans networks. From the simulation results, we
In this section, we present the comparative analysis of can conclude that the performance of the local and
proposed centrality measures for various real-world data global centralities is highly controlled by the network’s
sets. We have tested these measures namely GRACC,
14 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 9. Using the independent cascade model (IC model), the maximum information spread of the top − 10 influenced nodes of networks with varied
infection probabilities was normalized. Where D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local
and Global Centrality), ISC (Isolating Centrality), GRACC (Global Relative change in Average Clustering Coefficient), LRACC (Local Relative change in Average
Clustering Coefficient), GRAD (Global Relative change in Average Degree), LRAD (Local Relative change in Average Degree), GRAK (Global Relative change in
Average Katz), and LRAK (Local Relative change in Average Katz).

structure and properties. It has been also noted that GRACC, LRACC, GRAD, LRAD, GRAK, and LRAK
global centralities perform well compared with local outperformed other centralities such as D, B, C, CC, K,
centralities for networks with more number of cliques. PR, LGC, and ISC. We tested our centrality measures
on standard SIR and IC models to ensure maximum
VIII. CONCLUSIONS information spread. Kendall’s tau is used to determine
In this paper, we proposed six new centrality mea- whether the GRACC, LRACC, GRAD, LRAD, GRAK,
sures that make use of both local and global structural LRAK, and other existing centralities are equivalent.
information. First, we proposed generalized centrality Furthermore, we demonstrated that our proposed local
measures based on the relative change of the clustering centrality measures LRACC, LRAD, and LRAK re-
coefficient, degree, and Katz following node deletion. quire less computational time. Finally, we demonstrated
Then, we demonstrated that the proposed centralities that the proposed global centrality measures GRACC,
VOLUME x, 2022 15

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

GRAD, and GRAK outperform conventional measures [6] M. Ashtiani, A. Salehzadeh-Yazdi, Z. Razaghi-
in terms of information spread. One of the intriguing Moghadam, H. Hennig, O. Wolkenhauer, M.
future directions is to propose centrality measures that Mirzaie, and M. Jafari, “A systematic survey of
achieve maximum information spread while requiring centrality measures for protein-protein interaction
the least amount of computational time. One of our networks,” BMC systems biology, vol. 12, no. 1,
intuitions is that this can be achieved by combining pp. 1–17, 2018.
local and global centrality measures. Furthermore, the [7] S.-H. Li, D. C. Yen, W.-H. Lu, and C. Wang,
relative change of other centralities in the literature can “Identifying the signs of fraudulent accounts using
be generalized to investigate the efficacy of current ones. data mining techniques,” Computers in Human
Behavior, vol. 28, no. 3, pp. 1002–1013, 2012.
ABBREVIATIONS [8] M. Gosak, M. Milojević, M. Duh, K. Skok, and
The following are the abbreviations used in this paper: M. Perc, “Networks behind the morphology and
D: Degree centrality structural design of living systems,” Physics of
B: Betweenness centrality Life Reviews, vol. 41, pp. 1–21, 2022.
C: Closeness centrality [9] M. Jusup, P. Holme, K. Kanazawa, M. Takayasu,
K: Katz centrality I. Romić, Z. Wang, S. Geček, T. Lipić, B. Podob-
PR: PageRank nik, L. Wang, et al., “Social physics,” Physics
LGC: Local and global centrality Reports, vol. 948, pp. 1–148, 2022.
ISC: Isolating centrality [10] L. d. F. Costa, F. A. Rodrigues, and A. S. Cristino,
ASP: Average Shortest Path “Complex networks: The key to systems biology,”
GRASP: Global Relative change of Average Shortest Genetics and Molecular Biology, vol. 31, no. 3,
Path pp. 591–601, 2008.
LRASP: Local Relative change of Average Shortest [11] K. Stephenson and M. Zelen, “Rethinking cen-
Path trality: Methods and examples,” Social networks,
GRACC: Global Relative change in Average Cluster- vol. 11, no. 1, pp. 1–37, 1989.
ing Coefficient [12] F. Grando, D. Noble, and L. C. Lamb, “An anal-
LRACC: Local Relative change in Average Cluster- ysis of centrality measures for complex and social
ing Coefficient networks,” in Global Communications Conference,
GRAD: Global Relative change in Average Degree IEEE, 2016, pp. 1–6.
LRAD: Local Relative change in Average Degree [13] S. P. Borgatti, “Centrality and network flow,”
GRAK: Global Relative change in Average Katz Social networks, vol. 27, no. 1, pp. 55–71, 2005.
LRAK: Local Relative change in Average Katz [14] L. C. Freeman, “A set of measures of centrality
based on betweenness,” Sociometry, vol. 40, no. 1,
References pp. 35–41, 1977.
[1] D. C. Bell, J. S. Atkinson, and J. W. Carlson, [15] J. Zhan, S. Gurung, and S. P. K. Parsa, “Iden-
“Centrality measures for disease transmission net- tification of top-k nodes in large networks using
works,” Social networks, vol. 21, no. 1, pp. 1–21, katz centrality,” Journal of Big Data, vol. 4, no. 1,
1999. pp. 1–19, 2017.
[2] M. K. Enduri and S. Jolad, “Dynamics of dengue [16] W. Xing and A. Ghorbani, “Weighted pager-
disease with human and vector mobility,” Spatial ank algorithm,” in Second Annual Conference on
and spatio-temporal epidemiology, vol. 25, pp. 57– Communication Networks and Services Research,,
66, 2018. IEEE, 2004, pp. 305–314.
[3] K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, [17] F. Bloch, M. O. Jackson, and P. Tebaldi, “Cen-
“Fake news detection on social media: A data trality measures in networks,” arXiv preprint
mining perspective,” ACM SIGKDD explorations arXiv:1608.05845, 2016.
newsletter, vol. 19, no. 1, pp. 22–36, 2017. [18] J. Sheng, J. Dai, B. Wang, G. Duan, J. Long,
[4] J. Yang, C. Yao, W. Ma, and G. Chen, “A study of J. Zhang, K. Guan, S. Hu, L. Chen, and W.
the spreading scheme for viral marketing based on Guan, “Identifying influential nodes in complex
a complex network model,” Physica A: Statistical networks based on global and local structure,”
Mechanics and its Applications, vol. 389, no. 4, Physica A: Statistical Mechanics and its Applica-
pp. 859–870, 2010. tions, vol. 541, p. 123 262, 2020.
[5] G. Wang, Y. Wang, Y. Liu, and Y. Chi, “An [19] K. Berahmand, A. Bouyer, and N. Samadi, “A new
overview of structurally complex network-based centrality measure based on the negative and pos-
modeling of public opinion in the “we the media” itive effects of clustering coefficient for identifying
era,” International Journal of Modern Physics B, influential spreaders in complex networks,” Chaos,
vol. 32, no. 13, p. 1 830 006, 2018. Solitons & Fractals, vol. 110, pp. 41–54, 2018.
16 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

[20] K. L. Clark, U. Bhatia, E. A. Kodra, and A. R. [32] A. Ullah, B. Wang, J. Sheng, J. Long, N. Khan,
Ganguly, “Resilience of the us national airspace and Z. Sun, “Identifying vital nodes from local and
system airport network,” IEEE Transactions on global perspectives in complex networks,” Expert
Intelligent Transportation Systems, vol. 19, no. 12, Systems with Applications, vol. 186, p. 115 778,
pp. 3785–3794, 2018. 2021.
[21] T. Xu, J. Chen, Y. He, and D.-R. He, “Complex [33] O. Ugurlu, “Comparative analysis of centrality
network properties of chinese power grid,” Inter- measures for identifying critical nodes in com-
national Journal of Modern Physics B, vol. 18, plex networks,” Journal of Computational Science,
no. 17n19, pp. 2599–2603, 2004. vol. 62, p. 101 738, 2022.
[22] T. T. Sarı and G. Seçinti, “Centair: Central- [34] Y. Yang, X. Wang, Y. Chen, M. Hu, and C. Ruan,
ity based cross layer routing for software-defined “A novel centrality of influential nodes identifica-
aerial networks,” in 19th Annual Consumer Com- tion in complex networks,” IEEE Access, vol. 8,
munications & Networking Conference, IEEE, pp. 58 742–58 751, 2020.
2022, pp. 741–742. [35] G. Zhao, P. Jia, C. Huang, A. Zhou, and Y.
[23] D. T. Eliiyi, H. Arslan, V. K. Akram, and O. Fang, “A machine learning based framework for
Uğurlu, “Parallel identification of central nodes in identifying influential nodes in complex networks,”
wireless multi-hop networks,” in 28th Signal Pro- IEEE Access, vol. 8, pp. 65 462–65 471, 2020.
cessing and Communications Applications Confer- [36] K. Hajarathaiah, M. K. Enduri, and S. Anamala-
ence, IEEE, 2020, pp. 1–4. mudi, “Efficient algorithm for finding the influen-
[24] G. Yang, T. P. Benko, M. Cavaliere, J. Huang, tial nodes using local relative change of average
and M. Perc, “Identification of influential invaders shortest path,” Physica A: Statistical Mechanics
in evolutionary populations,” Scientific reports, and its Applications, vol. 591, p. 126 708, 2022,
vol. 9, no. 1, pp. 1–12, 2019. issn: 0378-4371.
[25] J. Sheng, J. Zhu, Y. Wang, B. Wang, and Z. Hou, [37] J. Zhong, F. Zhang, and Z. Li, “Identification of
“Identifying influential nodes of complex networks vital nodes in complex network via belief propa-
based on trust-value,” Algorithms, vol. 13, no. 11, gation and node reinsertion,” IEEE Access, vol. 6,
p. 280, 2020. pp. 29 200–29 210, 2018.
[26] K. Hajarathaiah, M. K. Enduri, S. Anamalamudi, [38] H. Zhang, S. Zhong, Y. Deng, and K. H. Cheong,
T. Subba Reddy, and S. Tokala, “Computing “Lfic: Identifying influential nodes in complex net-
influential nodes using the nearest neighborhood works by local fuzzy information centrality,” IEEE
trust value and pagerank in complex networks,” Transactions on Fuzzy Systems, vol. 30, no. 8,
Entropy, vol. 24, no. 5, p. 704, 2022. pp. 3284–3296, 2022.
[27] A. Zareie, A. Sheikhahmadi, M. Jalili, and [39] D. Baruah and A. Bharali, “A comparative study
M. S. K. Fasaei, “Finding influential nodes in of vertex deleted centrality measures,” Annals of
social networks based on neighborhood correlation Pure and Applied Mathematics, vol. 14, no. 1,
coefficient,” Knowledge-based systems, vol. 194, pp. 199–205, 2017.
p. 105 580, 2020. [40] L. C. Freeman, “Centrality in social networks
[28] X. Zhao, S. Xing, Q. Wang, et al., “Identifying conceptual clarification,” Social networks, vol. 1,
influential spreaders in social networks via nor- no. 3, pp. 215–239, 1978.
malized local structure attributes,” IEEE Access, [41] H.-L. Liu, C. Ma, B.-B. Xiang, M. Tang, and H.-F.
vol. 6, pp. 66 095–66 104, 2018. Zhang, “Identifying multiple influential spreaders
[29] Y. Xin, C. Gao, Z. Wang, X. Zhen, and X. Li, “Dis- based on generalized closeness centrality,” Phys-
cerning influential spreaders in complex networks ica A: Statistical Mechanics and its Applications,
by accounting the spreading heterogeneity of the vol. 492, pp. 2237–2248, 2018.
nodes,” IEEE Access, vol. 7, pp. 92 070–92 078, [42] D. J. Watts and S. H. Strogatz, “Collective dy-
2019. namics of ‘small-world’networks,” nature, vol. 393,
[30] P. De Lellis and M. Porfiri, “Detection of influen- no. 6684, pp. 440–442, 1998.
tial nodes in network dynamical systems from time [43] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez,
series,” IEEE Transactions on Control of Network and D.-U. Hwang, “Complex networks: Structure
Systems, vol. 8, no. 3, pp. 1249–1260, 2021. and dynamics,” Physics reports, vol. 424, no. 4-5,
[31] M. M. Tulu, R. Hou, and T. Younas, “Identifying pp. 175–308, 2006.
influential nodes based on community structure to [44] S. Brin and L. Page, “Reprint of: The anatomy
speed up the dissemination of information in com- of a large-scale hypertextual web search engine,”
plex network,” IEEE Access, vol. 6, pp. 7390–7401, Computer networks, vol. 56, no. 18, pp. 3825–3833,
2018. 2012.

VOLUME x, 2022 17

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

[45] A. N. Langville and C. D. Meyer, “Google’s KODURU HAJARATHAIAH received the B.

pagerank and beyond,” in Google’s PageRank and Tech degree in Computer science and en-
gineering from JNTUH, Hyderabad, India
Beyond, Princeton university press, 2011. in 2007 and M.Tech degree in computer sci-
[46] Z. Lv, N. Zhao, F. Xiong, and N. Chen, “A novel ence and engineering from JNTUA, Andhra
measure of identifying influential nodes in complex Pradesh, India in 2010. He is currently
networks,” Physica A: Statistical Mechanics and pursing a Ph.D. degree in computer Science
and Engineering from SRM University-AP,
its Applications, vol. 523, pp. 488–497, 2019. Amaravati, India. His research interests in-
[47] L. Tahmooresnejad and C. Beaudry, “The impor- clude Complex Networks and Analysis of
tance of collaborative networks in canadian scien- Algorithms, Information-Centric Networks (ICN) with IoT.
tific research,” Industry and Innovation, vol. 25,
no. 10, pp. 990–1029, 2018. MURALI KRISHNA ENDURI (Member
[48] J. Wang, H. Mo, F. Wang, and F. Jin, “Explor- IEEE) is an Assistant Professor at the
Department of Computer Science and Engi-
ing the network structure and nodal centrality neering at SRM University-AP, Amaravati,
of china’s air transport network: A complex net- India. He holds a M.Sc. in Mathematics
work approach,” Journal of Transport Geography, from the Acharya Nagarjuna University,
vol. 19, no. 4, pp. 712–721, 2011. Andhra Pradesh, India in 2008. He ob-
tained his M. Tech in Systems Analysis and
[49] R. A. Rossi and N. K. Ahmed, “The network data Computer Applications from the National
repository,” 2015. Institute of Technology Karnataka, India in
[50] R. Pastor-Satorras, C. Castellano, P. Van 2011. He received his Ph. D in Computer Science and Engineering
Mieghem, and A. Vespignani, “Epidemic processes from Indian Institute of Technology Gandhinagar, Gujarat, India
in 2018. He has a research experience/post Doc at Indian Institute
in complex networks,” Reviews of modern physics, of Technology Madras, Tamil Nādu, India in 2018. Murali Kr-
vol. 87, no. 3, p. 925, 2015. ishna Enduri’s research interest includes Algorithms, Complexity
[51] S. N. Dorogovtsev, A. V. Goltsev, and J. F. Theory and Complex Networks. Within these general areas, He
Mendes, “Critical phenomena in complex net- has a research experience on designing algorithms for graph
isomorphism problems for restricted graph classes, measuring the
works,” Reviews of Modern Physics, vol. 80, no. 4, impact of diversity of an article and predicting spread of disease.
p. 1275, 2008.
[52] D. Kempe, J. Kleinberg, and É. Tardos, “Maxi- SATEESHKRISHNA DHULI received the
mizing the spread of influence through a social net- B.Tech. degree in electronics and commu-
work,” in Proceedings of the ninth ACM SIGKDD nication engineering from the Maharaj Vi-
international conference on Knowledge discovery jayaram Gajapathi Raj College of Engi-
neering, Vizianagaram, India, in 2008, the
and data mining, 2003, pp. 137–146. M.Tech. degree in electronics and commu-
[53] M. Li, X. Wang, K. Gao, and S. Zhang, “A survey nication engineering from the National In-
on information diffusion in online social networks: stitute of Technology, Warangal, India, in
Models and methods,” Information, vol. 8, no. 4, 2011, and the Ph.D. degree in electrical
engineering from the Indian Institute of
p. 118, 2017. Technology, Kanpur, in 2019. He is currently working as an
[54] M. G. Kendall, “A new measure of rank corre- Assistant Professor with SRM University, Andhra Pradesh. His
lation,” Biometrika, vol. 30, no. 1/2, pp. 81–93, research interest includes wireless sensor networks, the Internet of
1938. Things, and communication systems.
[55] G. S. Shieh, “A weighted kendall’s tau statistic,”
Statistics & probability letters, vol. 39, no. 1, SATISH ANAMALAMUDI received the
B.Eng. degree in Computer Science and
pp. 17–24, 1998. Engineering from Jawaharlal Nehru Tech-
[56] Y. Liu, M. Tang, T. Zhou, and Y. Do, “Identify in- nological University, Hyderabad, India,
fluential spreaders in complex networks, the role of M.Tech in Network and Internet Engineer-
neighborhood,” Physica A: Statistical Mechanics ing from Karunya University, Coimbatore,
India and Ph.D. in Communication and In-
and its Applications, vol. 452, pp. 289–298, 2016. formation Systems from Dalian University
[57] H. Wei, Z. Pan, G. Hu, L. Zhang, H. Yang, X. Li, of Technology, Dalian, China. He worked
and X. Zhou, “Identifying influential nodes based as a Research Engineer in Beijing Huawei
on network representation learning in complex Technologies, Beijing, China, from November 2015 to Jan, 2017.
Currently, he is working as “Associate Professor” in “Department
networks,” PloS one, vol. 13, no. 7, 2018. of Computer Science and Engineering”, SRM University-AP, In-
[58] J. Zhu and L. Wang, “Identifying influential nodes dia. His research interests include common-control-channel design
in complex networks based on node itself and for MAC and routing protocols in cognitive radio ad hoc networks,
neighbor layer information,” Symmetry, vol. 13, MAC and routing protocol design of IoT and 5G networks.
no. 9, p. 1570, 2021.

18 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

LINGA REDDY CENKERAMADDI (Senior SUPPLEMENTARY INFORMATION

Member, IEEE) received the master’s de- In the supplementary information, we kept the results
gree in electrical engineering from the In-
dian Institute of Technology, New Delhi, of facebook data set. The facebook data set contains of
India, in 2004, and the Ph.D. degree in elec- 4039 nodes and 88234 edges.
trical engineering from the Norwegian Uni-
versity of Science and Technology (NTNU),
Trondheim, Norway, in 2011. Before join- A. DEGREE DISTRIBUTION
ing the Ph.D. degree, he worked at Texas The degree distribution plots for facebook is shown in
Instruments in mixed signal circuit design. Fig. 10. The degree distribution (neighbor distribution)
After finishing his Ph.D. degree, he worked in radiation imaging
for an atmosphere space interaction monitor (ASIM Mission is the most significant characteristic of a network struc-
to International Space Station) with the University of Bergen, ture. The x-axis is represented by degree, and the y-axis
Norway, from 2010 to 2012. He is currently the Group Leader of is represented by frequency.
the Autonomous and Cyber-Physical Systems (ACPS) Research
Group and working as a Professor with the University of Agder,
Grimstad Campus, Norway. He has coauthored over 90 research
publications that have been published in prestigious international
journals and standard conferences. His main scientific research
interests include cyber-physical systems, autonomous systems, and
wireless embedded systems. He is a member of the editorial boards
of various international journals, as well as the technical program
committees of several IEEE conferences. He is also a principal
(and co-principal) investigator of many research grants from the
Norwegian Research Council.

FIGURE 10. Degree distribution for facebook.

B. RANK CORRELATION OF GRACC, LRACC, GRAD,

LRAD, GRAK, AND LRAK WITH BASIC CENTRALITIES
The proposed centralities are not close to the basic
centralities such as D, B, C, CC, K, PR, LGC, and
ISC in the facebook network. The GRAD and GRAK
centralities are very much correlated with each other.
The LRACC and LRAD are correlated. Other central-
ities are not correlated in the facebook network, shown
in Fig. 11. From this subsection, we can observe that
the proposed centralities do not coincide with any basic
existing centralities.

C. SPREADING ABILITY WITH CENTRALITY VALUE

Considering the count of 100 simulations in the SIR
model, we study the relationship between the spread
ability and the centrality value of the node in this
section. Various centrality methods are assessed for the
centrality value of every node. The node with the highest
centrality value is considered as an infected node. With
100 times SIR model simulations, the total number of
infected people is calculated. The infection probability
β is set within the range 0.1 to 0.3. If the infection
probability goes beyond 0.4, then most of the people
in the network will be infected.

VOLUME x, 2022 19

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 13. 100 simulations of the SIR model’s cumulatively infected nodes
for facebook networks. The first 10 nodes listed are infected nodes, as
determined by proposed centralities and different centralities.
FIGURE 11. Rank Correlation with basic centralities for six data sets. Where
D (Degree), B (Betweenness), C (Closeness), CC (Clustering Coefficient), K
(Katz), PR (PageRank), LGC (Local and Global Centrality), ISC (Isolating
Centrality), GRACC (Global Relative change in Average Clustering is more. The GRACC centrality information is more in
Coefficient), LRACC (Local Relative change in Average Clustering
Coefficient), GRAD (Global Relative change in Average Degree), LRAD
the facebook network, shown in Fig. 14.
(Local Relative change in Average Degree), GRAK (Global Relative change in
Average Katz), and LRAK (Local Relative change in Average Katz).

FIGURE 12. Using SIR model, Centrality value with infection rate for six
networks in column wise. Where D (Degree), B (Betweenness), C (Closeness), FIGURE 14. Using the independent cascade model (IC model) with top − 10
CC (Clustering Coefficient), K (Katz), and GRACC (Global Relative change in seed nodes, we calculated the average information spread among GRACC,
Average Clustering Coefficient), LRACC (Local Relative change in Average LRACC, GRAD, LRAD, GRAK, and LRAK with other centralities for the
Clustering Coefficient), GRAD (Global Relative change in Average Degree), facebook network Where D (Degree), B (Betweenness), C (Closeness), CC
LRAD (Local Relative change in Average Degree), GRAK (Global Relative (Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local and Global
change in Average Katz), and LRAK (Local Relative change in Average Katz). Centrality), ISC (Isolating Centrality), GRACC (Global Relative change in
Average Clustering Coefficient), LRACC (Local Relative change in Average
Clustering Coefficient), GRAD (Global Relative change in Average Degree),
LRAD (Local Relative change in Average Degree), GRAK (Global Relative
D. CUMULATIVE INFECTED NODES FOR GRACC, change in Average Katz), and LRAK (Local Relative change in Average Katz).
LRACC, GRAD, LRAD, GRAK, AND LRAK WITH BASIC
CENTRALITIES
GRAD centrality spreading ability is greater when E. NORMALIZED MAXIMUM INFLUENCE FOR
compared with other centralities in the email-univ. In GRACC, LRACC, GRAD, LRAD, GRAK, AND LRAK
the facebook network, GRACC and GRAK centralities WITH BASIC CENTRALITIES
are at the top of the list for spreading information, This section displays the assessment of infection spread
shown in Fig. 13. We can find with the independent ability for the top-10 most influential nodes with differ-
cascade method with similar observations. Using the ent infection rate. These nodes are discovered by the
independent cascade model (IC model), we illustrate the D, B, C, CC, K, PR, LGC, ISC, GRACC, LRACC,
average number of nodes that are information retrieved GRAD, LRAD, GRAK, and LRAK centrality methods.
with different time scales in Fig. 14. Different centrality Evaluation of infection probability was estimated in
measurements are used to construct the seed nodes, the range of 0.1 and 0.5 by using the SIR model (100
which are input to IC model. 1000 iterations were em- Simulations). In the spreading information point in the
ployed for simulations in the IC model. In the facebook facebook network, GRAK and GRACC centralities are
network, GRAK centrality average information spread at the highest position, shown in the Fig. 15.
20 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3232288

FIGURE 15. Using the SIR simulations, the maximum information spread of
the top − 10 influenced nodes of networks with varied infection probabilities
was normalized. Where D (Degree), B (Betweenness), C (Closeness), CC
(Clustering Coefficient), K (Katz), PR (PageRank), LGC (Local and Global
Centrality), ISC (Isolating Centrality), GRACC (Global Relative change in
Average Cluster Coefficient), LRACC (Local Relative change in Average
Clustering Coefficient), GRAD (Global Relative change in Average Degree),
LRAD (Local Relative change in Average Degree), GRAK (Global Relative
change in Average Katz), and LRAK (Local Relative change in Average Katz).

FIGURE 16. Using the independent cascade model (IC model), the
maximum information spread of the top − 10 influenced nodes of networks
with varied infection probabilities was normalized. Where D (Degree), B
(Betweenness), C (Closeness), CC (Cluster Coefficient), K (Katz), PR
(PageRank), LGC (Local and Global Centrality), ISC (Isolating Centrality),
GRACC (Global Relative change in Average Clustering Coefficient), LRACC
(Local Relative change in Average Clustering Coefficient), GRAD (Global
Relative change in Average Degree), LRAD (Local Relative change in Average
Degree), GRAK (Global Relative change in Average Katz), and LRAK (Local
Relative change in Average Katz).

In Fig. 16, using IC model, we describe maximum

information spread with varied probability infection.
The GRACC centrality is leading position in facebook
network, shown in Fig. 16. Similar observations are
clearly seen from both the methods of SIR and IC.
All the codes and data sets available in fol-
lowing link: https://github.com/endurimuralikrishna/
influential-nodes.

VOLUME x, 2022 21

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/