Attributes in International Trade

Accepted Manuscript
The use of nodes attributes in social network analysis with an application

to an international trade network
Ricardo Lopes de Andrade, Leandro Chaves Rêgo
PII: S0378-4371(17)30857-9
DOI: http://dx.doi.org/10.1016/j.physa.2017.08.126
Reference: PHYSA 18572
To appear in: Physica A
Received date : 8 November 2016

Revised date : 16 June 2017
Please cite this article as: R.L. de Andrade, L.C. Rêgo, The use of nodes attributes in social
network analysis with an application to an international trade network, Physica A (2017),
http://dx.doi.org/10.1016/j.physa.2017.08.126
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
The Use of Nodes Attributes in Social Network Analysis
with an Application to an International Trade Network
Ricardo Lopes de Andradea,∗, Leandro Chaves Rêgob
a
Universidade Federal de Pernambuco. Programa de Pós-Graduação em Engenharia de
Produção, Avenida da Arquitetura, CEP: 50.740-550, Recife-PE, Brazil.
b
Universidade Federal do Ceará. Departamento de Estatı́stica e Matemática Aplicada,
Rua Campus do Pici, CEP: 60440-554, Fortaleza-CE, Brazil.
Abstract
The social network analysis (SNA) studies the interactions among actors in
a network formed through some relationship (friendship, cooperation, trade,
among others.). The SNA is constantly approached from a binary point of
view, i.e., it is only observed if a link between two actors is present or not re-
gardless of the strength of this link. It is known that different information can
be obtained in weighted and unweighted networks and that the information
extracted from weighted networks is more accurate and detailed. Another
rarely discussed approach in the SNA is related to the individual attributes
of the actors (nodes), because such analysis is usually focused on the topolog-
ical structure of networks. Features of the nodes are not incorporated in the
SNA what implies that there is some loss or misperception of information in
those analyze. This paper aims at exploring more precisely the complexities
of a social network, initially developing a method that inserts the individual
attributes in the topological structure of the network and then analyzing the
network in four different ways: unweighted, edge-weighted and two methods
for using both edge-weights and nodes’ attributes. The international trade
network was chosen in the application of this approach, where the nodes
represent the countries, the links represent the cash flow in the trade trans-
actions and countries’ GDP were chosen as nodes’ attributes. As a result, it
is possible to observe which countries are most connected in the world econ-
∗
Corresponding author at:
Email addresses: ricardolopesa@gmail.com (Ricardo Lopes de Andrade),
leandro@dema.ufc.br (Leandro Chaves Rêgo)
Preprint submitted to Physica A September 14, 2017

omy and with higher cash flows, to point out the countries that are central
to the intermediation of the wealth flow and those that are most benefited
from being included in this network. We also made a principal component
analysis to study which metrics are more influential in describing the data
variability, which turn out to be mostly the weighted metrics which include
the nodes’ attributes.
Keywords: Nodes’ Attributes, Weighted Networks, International Trade
Network
1. Introduction
A social network can be defined as a set of people or actors, where each
one of them has some sort of connection (link) to some or all other actors [1].
Both actors and links can be established in different ways, depending on
the context. An actor can be a single person, a group, a company or even
countries. The link can be a friendship relation between two people, a collab-
oration or the existence of a joint member between two teams, or a business
relationship among companies [2], as well as trade relations between coun-
tries.
Recently, the study of the relationships present in social networks are
receiving great attention with the aim of characterizing their topological
complexities, namely: the analysis of their links’ patterns, interactions, and
implications. Such analysis has been facilitated through the use of Graph
Theory and a set of mathematical algorithms [3]. This approach is known as
Social Network Analysis (SNA).
The networks, besides having a complex structure, present a great het-
erogeneity in the capacity and intensity of links [4]. This difference between
links can be assessed using a weighted network, in which it is considered not
only the presence or absence of a link between two nodes (binary relation),
but some kind of weight that captures the strength of the connection [5].
Thus, in a weighted social network there may be stronger or weaker links
between nodes. For Granovetter [6], the link’s weight can be defined by
the frequency of the relationship, emotional intensity, intimacy and mutual
service between nodes.
It is observed that many SNA studies are based on unweighted networks,
as most SNA measures are only defined for binary situations, and thus unable
to cope with weighted networks [7]. Thus, when analyzing a weighted network
2
in an unweighted manner, most of the information contained in the data set
is lost, and consequently, the topological complexity of the network cannot
be described in its full extent or richness [5, 7].
The unweighted and weighted networks provide different information [8].
Therefore, studying the networks only from a single perspective, unweighted
or weighted, can generate inaccurate results. Whenever possible, the ideal
analysis should combine both weighted and unweighted metrics, as in the
study by Schiavo et al. [9] in an international trade network. According to
them, analyzing both weighted and unweighted networks can deliver more
accurate and detailed insights about the topological structure and, conse-
quently, about international trade properties.
Another feature that has been discussed in SNA is the lack of inclusion of
the individual attributes of the nodes [10] while analyzing a social network.
Thus, by focusing only in the topological structure, the characteristics of
the nodes are ignored [11]. However, these features (which define the node’s
attribute) may influence the network structure, revealing new standards, new
interactions and, consequently, new implications.
In an attempt to explore with greater extent the complexities of a social
network, this paper aims to develop a method that inserts the node’s at-
tributes in the topological structure of the network. This proposed method
is applied in a social network whose relations are established by commer-
cial transactions among countries, an international trade network. The SNA
metrics of such network were calculated in four different ways: unweighted,
weighted and two methods of including node’s attributes in a weighted net-
work. A comparison of the results obtained was made. Through the analysis
of the values of these metrics, it can, for example, be observed which countries
are most connected in the world economy and with higher cash flows, how
close are the relations among them, find out the countries that are central to
the intermediation of the wealth flow and those that are most benefited from
being included in this network. These results, besides showing the useful-
ness of adding the node’s attributes in the SNA analysis, can help to better
understand the global international trade flows.
The remaining part of this work is divided as follows: Section 2 presents
the SNA metrics adopted in this work; Section 3 presents the method that
inserts the node’s attributes in the network metrics; the application of the
proposed method in an international trade network among countries is pre-
sented in Section 4, where we also analyze which metrics are more important
in describing data variability through a principal component analysis. Fi-
3
nally, Section 5 presents the last considerations of the study and proposals
for future work.
2. Metrics
Mathematically, a graph, G, is defined as a structure constituted by a
finite and not empty set of nodes (vertices), V (G) = {v1 , v2 , · · · , vn }, and a
finite set of edges (links) E(G), consisting of pairs of nodes from V (G) (not
necessarily ordered or different).
The order or cardinality of a graph is the number of elements of the set
of vertices V (G), i.e., #V (G). The edge set E(G) leads to a binary relation
in V (G) and is called an adjacency relationship between the vertices of G.
A network with n vertices is represented by an adjacency matrix A(G) with
n × n elements, where

1, if vi and vj are connected,
a(vi , vj ) =
0, otherwise.
If an incident edge connects vertices vi and vj of a graph, then it is said
that these vertices are neighbors or adjacent. The neighborhood of vertex vi
is the set formed by all vertices adjacent to it.
If in a graph each edge has a non-negative value or associated weight,
this graph is called weighted and can be represented by a weighted adjacency
matrix, W (G), whose inputs are not simply 0 or 1, but are equal to weights of
the edges [7]: w(vi , vj )=(weight of the link between vi and vj ). If w(vi , vj ) =
w(vj , vi ) for all vi and vj in V (G), then G is called an undirected graph.
Otherwise, it is called a directed graph.
A path is a sequence of vertices and edges, where each edge connects the
vertex that precedes it to the vertex that follows it, not repeating edges. The
length of a path between two given vertices will be given by the number of
edges present in the path.
A geodesic path or shortest path is the shortest path between two vertices
[7]. The geodesic path length, d(vi , vj ), also called geodesic distance or short-
est distance, thus, is the shortest distance in the network between these two
vertices. Formally, given a path c = (v0 , v1 , v2 , · · · , vk ) between the vertices
vi and vj , the length of this path is given by dc = k. Let C(vi , vj ) be the set
of all paths between vertices vi and vj . Thus, the geodesic distance is defined
by:
4
d(vi , vj ) = min{dc : c ∈ C(vi , vj )}. (1)
In the case of weighted networks, the length of a path c = (v0 , v1 , v2 , · · · vk )
between vertices vi and vj , can be formally defined by Dijkstras algorithm
Newman [1] and Brandes [12]:
1 1 1
dw
c = w(v0 ,v1 )
+ w(v1 ,v2 )
+ · · · w(vk−1 ,vk )
And the weighted geodesic distance is given by:
dw (vi , vj ) = min{dw
c : c ∈ C(vi , vj )}. (2)
The identification of the shortest paths and their length in directed net-
works is similar to the process in undirected networks. The unique difference
is that a path from one vertex to another can only follow the direction of the
edges.
The degree centrality, proposed by Freeman [13], is calculated in terms
of the number of adjacent vertices, i.e., the degree centrality of vertex vi ,
denoted by Cd (vi ), is the number of vertices adjacent to vertex vi . Formally,
the degree centrality is defined by:
n
X
Cd (vi ) = a(vi , vj ). (3)
j=1
If the network is weighted, then the weighted degree centrality of vertex

vi is equal to the sum of the weights of the edges that are connected to the
vertex vi . For Newman [7] and Barrat et al. [4] the weighted degree centrality
is defined by:
n
X
Cdw (vi ) = w(vi , vj ). (4)
j=1
For weighted directed networks, degree centrality has two forms: in-degree
centrality and out-degree centrality. The in-degree of vertex vi is the sum
of weights of the links destined to vi , denoted by din
i , while the out-degree
of vertex vi is the sum of the weights of the links originated from vi and
is denoted by dout
i . When we combine in-degree and out-degree, we are in
essence disregarding vertices directions. Thus, Cdw (vi ) = din out
i + di .
Another metric to analyze a node on the network started from the theory
of “strong” links of Krackhardt [14]. The average link strength of vertex vi is
5
defined as the ratio between the weighted degree centrality, Cdw (vi ), and the
degree centrality, Cd (vi ):
Cdw (vi )
LS(i) = . (5)
Cd (vi )
Therefore, LS(i) represents the average link weight of node vi . A metric
that takes into consideration the geodesic distance from a given node to all
other nodes of the network is the closeness centrality. Freeman [13] asserted
that the closeness centrality of vertex vi , denoted by Cc (vi ), is given by:
1
Cc (vi ) = P . (6)
j d(vi , vj )
The most central vertices in a network according to this metric are those
that have a smaller distance to the other vertices. In weighted networks, the
weighted closeness centrality is given by:
1
Ccw (vi ) = P w
. (7)
j d (vi , vj )
The betweenness centrality of vertex vi is the sum, for every pair of nodes
different from vi , of the ratio between the number of shortest paths between
the given pair of nodes that go through vi , and the total number of shortest
paths between the given pair of nodes [3, 13]. The betweenness centrality,
Cb (vi ), of vertex vi is given by:
X g(vj , vi , vk )
Cb (vi ) = , j 6= k 6= i, (8)
j,k
g(vj , vk )
where g(vj , vk ) is the number of shortest paths between vertex vj and vertex
vk and g(vj , vi , vk ) is the number of shortest paths between vertex vj and
vertex vk going through vi .
In a weighted network, the betweenness centrality is given by:
X g w (vj , vi , vk )
Cbw (vi ) = , j 6= k 6= i, (9)
j,k
g w (vj , vk )
where g w (vj , vk ) is the number of weighted shortest paths between vertex

vj and vertex vk and g w (vj , vi , vk ) is the number of weighted shortest paths
6
between vertex vj and vertex vk going through vi , considering the weighted
distance, dw (vi , vj ).
A metric of importance of a vertex in the network based on its connec-
tions, the eigenvector centrality is supported on the idea that a particular
node will have high centrality if it is connected to vertices with central po-
sitions in the network [15]. In other words, the centrality of the vertex does
not depend only on the number of vertices adjacent to it, but also on the cen-
trality of these vertices. Let λ be a constant, then the eigenvector centrality
of Ce (vi ) is given by:
n
1X
Ce (vi ) = a(vi , vj )Ce (vj ). (10)
λ j=1
Using the vector notation, let X = (Ce (1), Ce (2), · · · , Ce (n)) be the vec-
tor of eigenvector centralities, we can rewrite Equation (10) as λX = AX.
By assuming that the eigenvector centrality assumes only non-negative val-
ues (using the Perron-Frobenius theorem), it can be shown that λ is the
largest eigenvalue of the adjacency matrix, where X is the corresponding
eigenvector [16].
In the case of weighted networks, the elements of the adjacency matrix
are the weights of the edges, w(vi , vj ), [7]. And the eigenvector centrality is
defined by:
n
1X
Cew (vi ) = w(vi , vj )Cew (vj ). (11)
λ j=1
For directed networks, the eigenvector centrality of the vertex vi is defined
considering the edges destined to vi . Therefore, the centrality vector X is the
left-hand eigenvector of the weighted adjacency matrix W associated with
the largest eigenvalue λ.
The local clustering coefficient indicates how the adjacent nodes of a given
node are connected to each other and, together with the average value of the
shortest path, it can identify a “small-world” effect (networks with large
cluster coefficient and relatively short distance between nodes) [17]. The
clustering coefficient of a vertex vi is the ratio between the number of trian-
gles that contains vertex vi and the number of possible edges between the
neighboring vertices. Let N T (vi ) be the number of triangles (consists of three
nodes connected by three links) containing vertex vi . For Onnela et al. [18],
the local cluster coefficient is defined as:
7
2N T (vi )
CCL(vi ) = . (12)
Cd (vi )(Cd (vi ) − 1)
The weighted local clustering coefficient was proposed by Onnela et al.
[18] and is given by:
2 X
CCLw (vi ) = ((ŵ(vi , vj )ŵ(vi , vk )ŵ(vj , vk ))1/3 , (13)
Cd (vi )(Cd (vi ) − 1) j,k
where the weights of the edges are normalized by the maximum weight of the
network, ŵ(vi , vj ) = w(vi , vj )/maxvi ,vj ∈V (G) (w(vi , vj )) and the contribution of
each triangle depends on all the weights of the edges.
The PageRank is a method of ranking web pages, measuring effectively
the interest of browsers and attention devoted to them [19]. The PageRank
considers the number and quality of links to a web page in order to determine
how influential it is [11]. Let TA be a web page and Ti be one of the web pages
that connects to TA . Brin and Page [20] defined the PageRank as follows:

P R(T1 ) P R(Tn )
P R(TA ) = (1 − δ) + δ + ··· + , (14)
C(T1 ) C(Tn )
where P R(TA ) is the PageRank of page TA , P R(Ti ) is the PageRank of page
Ti , C(Ti ) is the number of outbound links on page Ti and δ is a damping
factor (assuming that a person randomly clicks on pages and eventually stops
clicking, δ is the probability at any given moment, the person will continue
to click), which can be set between 0 and 1. This setting is applied on both
directed and undirected networks. In this paper, we adopted the value of
δ = 0.85.
In the study by Santos [21] on co-authorship networks, it was proposed a
metric to evaluate the benefit or utility for a given author of belonging to a
certain network structure. According to this metric, it is considered that each
author has a finite amount of time to devote to scientific collaborations, and
that each author receives a utility from an adjacent author who is equal to
the proportion of papers that the co-author has with him more the formation
of a synergy, which is given by the product of the dedication of each author
to the collaboration. Formally, a utility U w (vi ) of a given author vi in a given
graph G is given by:
8
X w(vi , vj ) w(vi , vj ) w(vi , vj )2

w
U (vi ) = + w + w , (15)
j
Cdw (vi ) Cd (vj ) Cd (vi )Cdw (vj )
where w(vi , vj ) is the weight of the link between authors vi and vj , Cdw (vi )
and Cdw (vj ) are the weighted degrees of these authors, respectively.
The utility developed by Santos [21] was based on the original model
of the utility of Jackson and Wolinsky [22]. This model takes into account
only if the author is or is not connected to another author, disregarding the
weights of the links. Thus the unweighted utility of a particular author vi in
a given graph G is given by:
X 1 1 1

U (vi ) = + + , (16)
j
C d (vi ) C d (vj ) C d (v i )C d (vj )
where Cd (vi ) and Cd (vj ) are the centrality degree of vertices vi and vj , re-
spectively.
For a directed network, there is no definition of utility in the literature.
To overcome this problem, we propose the following definition for utility in
directed networks. Let V in (vi ), V out (vi ) be the set of neighbors of vi which
have a connection arriving and leaving vi , respectively. Thus, dout i is the
cardinality of V out (vi ). Then, the utility of node vi is defined by:
X X
U (vi ) = (w(vj , vi )/dout
j )+ (w(vi , vj )/dout
i )+
j∈V in (vi ) j∈V out (vi )
X (17)
+ ((w(vi , vj )w(vj , vi )/dout out
i dj ))
j∈V in (vi )∩V out (vi )
It is worth pointing out that if w(vi , vj ) = w(vj , vi ) for all i and j in the
above formula, we have the same value for the utility as that for undirected
networks, given by 15.
The efficiency centrality1 is a new metric proposed by Wang et al. [23].
This metric is based on the concept of network efficiency by Latora and
Marchiori [24], expressed as EF (G), which measures the efficiency of network
information exchange. For the authors, the efficiency of the communication
1
We would like to thank an anonymous referee for pointing out this recently published
centrality measure.
9
between two vertices vi and vj can be defined by the inverse of the geodesic
distance that separates them, ij = 1/dij . And when there is no path that
connects the vertices vi and vj , dij = +∞ and, consequently, ij = 0. Then,
to measure the average flow of information on G, we calculate the average of
ij , which represents the efficiency EF (G) [24].
P
i6=j∈G ij 1 X 1
EF (G) = = (18)
N (N − 1) N (N − 1) i6=j∈G dij
Then, Wang et al. [23] used a measure to quantify the importance of a

single node in the graph G. When removing a node k from G, the graph
G reduces to a subgraph G0k , which has efficiency EF (G0k ). The efficiency
centrality CkEf f C of a node k is defined by the relative change in network
efficiency caused by the removal of node k from the initial graph G [23]:
∆EF EF [G] − EF [G0k ]

CkEf f C = = , k = 1, . . . , N. (19)
EF EF [G]
We think it is appropriate to comment on this measure. By definition,
when removing a node k from the graph G, where k is an isolated node in
the graph, such node would have efficiency centrality equal to ( N−2 −2
). The
fact that the efficiency of isolated nodes is negative is difficult to interpret.
This problem occurs because the subgraph G0k has one node less than G. In
order to avoid that problem, instead of removing node k from graph G, we
suggest to eliminate only its connections, if it is not isolated. In that way,
both the original graph and the subgraph have the same number of nodes.
With such modification, the efficiency centrality of an isolated node would be
zero. Such modification does not change the centrality ranking of the nodes,
since it is just a scaling change that is being realized. In this paper, we use
such modified efficiency centrality measure to avoid negative values.
In case of weighted networks, the efficiency of the communication between
two vertices vi and vj can be defined by the inverse of the weighted geodesic
distance, w w
ij = 1/dij (this distance is shown in Equation (2)). For a directed
network, the unique difference is that the path from one vertex to another
can only follow the direction of the edges. Then we should not only observe
the distance from vi to vj , but also from vj to vi .
10
3. Adding the Node’s Attributes in the Metrics
The main focus of the study of social networks is to analyze how the
relationships that individuals establish among themselves (the social struc-
ture) influences behavior and their performance [25], not worrying about the
individual attributes of the same [10]. Thus, the network-based methods
focus only on the topological structure, ignoring the characteristics of the
nodes [11]. However, these characteristics can influence and be influenced by
the social structure.
Liu et al. [11], in a study of co-authorship networks, developed a method
that combined the importance of nodes with the topological structure of the
network. Thus, in that work, the position or status of a node in the network
depends not only the structure, but also on the individual importance of the
node. Let the effective distance, D(vi , vj ), between two vertices, vi and vj ,
be given by:
D(vi , vj ) = 1 − log(w(vi , vj )/Cdw (vi )), (20)

where Cdw (vi ) is the weighted degree centrality of vertex vi . Liu et al. [11]
inserted the importance of the nodes modifying the weight of the edges by
means of Equation (21), which has similarity to the gravity law.
ks(vi )s(vj )
S(vi , vj ) = , (21)
D2 (vi , vj )
where S(vi , vj ) is the new weight of the edge connecting the vertices vi and
vj considering the node’s attributes, s(vi ) and s(vj ), and k is a constant. The
node’s attribute is defined by the number of weighted citations by means of
an equation developed by the authors that takes into consideration how old
the publication is.
As D(vi , vj ) and D(vj , vi ) may be different, there may be two different
weights between vertices vi and vj . Thus, their method may transform an
undirected into a directed network.
In this sense, seeking also to incorporate some information about the
nodes in the topological structure of a social network, we propose a new
method of redefining the weights that does not transform an undirected net-
work into a directed one, as follows:

s(vi ) + s(vj )
Z(vi , vj ) = w(vi , vj ) , (22)
2
11
where Z(vi , vj ) is equal to the original weight, w(vi , vj ), of the edge between
the vertices, vi and vj , multiplied by the average of both nodes’ attributes.
The node’s attributes are measurable characteristics associated with the type
of relationship that connects them. With the incorporation of the node’s
attributes in the network, Z(G) shall be the updated weighted adjacency
matrix, where Z(vi , vj ) is an element of this matrix. We assume that the
node’s attribute is a positive real number.
The similarity of the model proposed in this work with the model devel-
oped by Liu et al. [11] is limited to the fact that both seek to incorporate
information about the nodes in the network structure. In our model, when
the nodes’ attributes are not considered, it is assumed that they are all equal
to one, and therefore, the updated weighted adjacency matrix is a generaliza-
tion of the weighted adjacency matrix in which information about the nodes’
attributes is used.
On the other hand, in the model developed by Liu et al. [11], even if all
nodes have the same value of the attribute, the edges’ weights are modified
according to Equation (21). Although, in some applications this should make
sense, we believe it is important to distinguish the effects caused by the
differences in nodes’ status from the differences in link strengths. Moreover,
in Liu et al.’s method, the importance of node’s attributes is overemphasized
when compared to the importance of the edges’ weights, since the nodes’
attributes are multiplied and the edge weight is inside a logarithm.
It is observed that, unlike the model proposed by Liu et al. [11], Z(G)
is symmetric if and only if W (G) is. So, when the proposed method is
applied in an undirected network, it remains undirected. Since we do not
view any reason for why the inclusion of node’s attributes should transform
a symmetric relation between nodes into an asymmetric one, we view that
it is important to seek other methods of incorporating such attributes into
SNA.
In the SNA literature, we usually find the analysis of undirected networks,
since the study of the properties of directed networks, specifically in the
weighted case, may become more complex and, often, lead to less precise
results [26]. Having said that, we emphasize that the proposed method can
be applied to any type of network, either directed or undirected, and does
not change its nature.
The choice of which node’s attribute to use is at the analyst’s criterion
that seeks to analyze the changes occurring in the SNA with the insertion of
those attributes. Certain attributes may favor some nodes, relative position
12
in the network, while others may impair and vice versa. It is worth pointing
out that while performing a SNA, not only the choice of the node’s attribute
impact the results, but also the choice of what is defined to be a link between
nodes and both of these should be at the analyst’s criterion of what he is
interested in studying.
In a co-authorship or citation networks, for example, we can use the
performance indicators of the articles as node’s attributes. In studies of
networks that analyze the formations of couples, [27], the node’s attributes
may be linked, for example, to cultural or economic factors. In networks that
analyze disease propagations, node’s attributes may be related to risk factors.
For networks analyzing the diffusion of innovation, [28], node’s attributes can
be linked to the degree of innovation, the entrepreneurial profile, power of
capital, access to information, among other characteristics.
4. Application
To illustrate the results of the SNA metrics calculated in four distinct
ways (unweighted, edge-weighted and two methods of including node’s at-
tributes in edge-weighted networks), we used data of the International Trade
Center (ITC) (site: http://www.intracen.org/) for the formation of the in-
ternational trade network. With the list of exported and imported values of
178 countries in 2015, we have established the relations, the weight of the
edge between two countries was considered to be the average of the values
exported and imported between two countries (as in the database, the im-
ported value from a country A in relation to a Country B, sometimes was
different from the exported values from a country B to a country A, we de-
fine the average of these values as the approximate value of this data). The
gross domestic product (GDP) was defined as the node’s attribute. The net-
work is made up of 178 nodes and 10,419 edges, forming a single connected
component.
The following tables present the classification of the top 20 countries in
the network according to the SNA metrics calculated in four distinct ways:
unweighted, weighted and weighted including the node’s attributes according
to our proposed method and also according to Liu et al.’s method. For Liu et
al.’s method, we use k = 1 in Equation 21. Furthermore, we present scatter
plots between the results of two out of four different methods of calculation
of the metrics and also tables containing the Kendall correlations between
13
the results of different methods. The logarithmic scale was used in all scatter
plots that follow.
4.1. Degree Centrality

Table 1 ranks the top 20 countries according to the degree centrality,
calculated in four distinct ways: unweighted degree centrality - UDC; W-
weighted degree centrality - WDC; Z-weighted degree centrality - ZDC; and
L-weighted degree centrality - LDC. The four measures point to different as-
pects. The UDC highlights the number of trade links established by the coun-
try, disregarding the weight of the link (money flow). The WDC highlights
the financial value of the business transaction and disregards the number of
trading partners. Therefore, countries with smaller trade relations and high
volume of transactions and countries with higher trade relations, but with
low volume of transactions, will be seen in a similar manner. Finally, the
ZDC and LDC modify the money flow by incorporating the node’s attributes
which, in this particular network, is defined to be the node’s GDP, represent-
ing the wealth or the economic importance of the country. According to this
metric, an edge with smaller weight attached to a country with high GDP
or an edge with higher weight attached to a country with low GDP will be
accounted similarly.
The four degree centrality measures are analyzed in parallel. Table 1
shows how the number of links, the weights of the links and the combination
of the weights of the links with the node’s attributes change the countries’
ranking.
Note that the 16 most central countries, according to UDC, have the
same unweighted degree centrality. New Zealand, Denmark and Ireland also
have UDC equal to 176 and are, together with the other countries shown in
the table with the same degree centrality, the second set of countries with
the highest unweighted degree centrality. The United States is the most
central country, regardless of the use of the edges weights or of the node’s
attributes. Thus, this country is linked to all the other countries and presents
the highest money flow. China and Germany have the same number of links,
however, Germany presents a higher money flow and, therefore, became a
position above China when considering the edges weights. When the node’s
attributes are introduced, according to ZDC, the positions between these
two countries are reversed. Since both are connected to the same countries,
in order to China overtake Germany the values transacted by China with
countries of higher GDP were higher than those transacted by Germany.
14
UDC WDC ZDC LDC
Rank Country Value Country Value Country Value Country Value
(×109 ) (×1015 ) (×1015 )
1 United 177 United 16.63 United 37.71 United 73.64
States States States States
2 China 177 Germany 14.40 China 27.31 China 54.30
3 Japan 177 China 13.14 Germany 7.539 Japan 14.72
4 Germany 177 France 7.881 Canada 6.135 Mexico 12.18
5 United 177 Hong 6.679 Japan 6.075 Germany 11.27
Kingdom Kong
6 France 177 Canada 6.593 Mexico 5.536 United 8.508
Kingdom
7 India 177 Japan 6.058 United 3.622 France 6.318
Kingdom
8 Italy 177 Netherlands 5.530 Korea 3.423 India 5.868
9 Korea 177 United 5.508 Hong 3.332 Canada 5.856
Kingdom Kong
10 Australia 177 Italy 4.579 France 2.931 Korea 5.321
11 Spain 177 Switzerland 3.657 Italy 2.066 Italy 4.828
12 Netherlands 177 India 3.291 Netherlands 2.008 Brazil 4.781
13 Switzerland 177 Belgium 3.043 India 1.827 Australia 4.192
14 Thailand 177 Korea 3.042 Australia 1.339 Indonesia 3.982
15 Malaysia 177 Poland 2.806 Brazil 1.331 Russia 3.162
16 Singapore 177 Russia 2.705 Belgium 1.307 Spain 2.955
17 Mexico 176 Spain 2.444 Switzerland 1.266 Iran 2.071
18 Sweden 176 Austria 2.268 Singapore 1.237 Venezuela 2.060
19 Austria 176 Singapore 2.228 Russia 1.148 Netherlands 2.058
20 South 176 Brazil 2.018 Vietnam 1.117 Nigeria 1.755
Africa
Table 1: The top 20 countries according to the degree centrality
According to LDC, Germany loses two positions, while Japan and Mexico go
up two. It is worth pointing out that the inclusion of the node’s attributes,
in both methods, highlighted the importance of China in the network since
it got a much bigger advantage to the country in third place, as opposed to
the case where WDC was used where it got a result close to Germany.
Figure 1 shows scatter plots for the degree centrality metrics. We can see
that there is a positive relationship between the metrics. The metric with
highest correlation with UDC is WDC. Regarding the relation between UDC
and ZDC and between UDC and LDC, the formation of groups is observed.
The rationale for this effect is the fact that countries with lower trade and
high volume of transactions and countries with greater trade, but with a low
volume of transactions will be seen in a equivalent manner according to ZDC
or LDC. However, in the UDC-LDC association, one group is isolated, dis-
torting the relationship between the variables. This fact also occurs in other
relationships when LDC is one of the variables. In general, the WDC, ZDC,
and LDC metrics can differentiate countries that get the same UDC value.
Therefore, the inclusion of more information in the network contributes pos-
itively to the distinction of countries. Table 2 shows the correlation between
these metrics, as expected the LDC metric reveals the lowest correlations
with UDC and WDC. On the other hand, LDC showed the highest correla-
15
tion with ZDC, what shows a kind of robustness of the way node’s attributes
affects the values of the metrics in both methods.
UDC
WDC
ZDC
LDC
UDC WDC ZDC LDC
Figure 1: Relations between the degree centrality metrics
UDC WDC ZDC LDC

UDC 1 - - -
WDC 0.751 1 - -
ZDC 0.579 0.761 1 -
LDC 0.501 0.645 0.817 1
Table 2: Correlations between the degree centrality metrics

Page 1
4.2. Average Link Strength

The result of the top 20 countries according to the average link strength is
shown in Table 3. The W-average link strength (WLS) is the ratio between
WDC and UDC; Z-average link strength (ZLS) is the ratio between ZDC
and UDC; and L-average link strength (LLS) is the ratio between LDC and
LDC. The United States, Germany and China have the strongest links in
WLS and ZLS . According to LLS, Germany loses two positions, while Japan
and Mexico go up two.
Figure 2 shows the relationship among the WLS, ZLS and LLS. Here we
also observe the formation of a dissonant group of countries when correlated
with the LLS metric. Table 4 shows the correlation between these metrics.
16
WLS ZLS LLS
Rank Country Value Country Value Country Value
(×107 ) (×1012 ) (×1013 )
1 United 9.398 United 213.1 United 41.60
States States States
2 Germany 8.135 China 154.3 China 30.68
3 China 7.425 Germany 42.59 Japan 8.314
4 France 4.452 Canada 35.46 Mexico 6.923
5 Hong 3.861 Japan 34.32 Germany 6.367
Kong
6 Canada 3.811 Mexico 31.45 United 4.807
Kingdom
7 Japan 3.423 United 20.46 Indonesia 4.425
Kingdom
8 Netherlands 3.124 Korea 19.34 France 3.569
9 United 3.112 Hong 19.26 Canada 3.385
Kingdom Kong
10 Italy 2.587 France 16.56 India 3.315
11 Switzerland 2.066 Vietnam 12.41 Korea 3.006
12 India 1.859 Italy 11.67 Brazil 2.732
13 Belgium 1.739 Netherlands 11.34 Italy 2.728
14 Korea 1.719 India 10.32 Australia 2.368
15 Poland 1.631 Indonesia 9.120 Venezuela 2.368
16 Russia 1.573 United 8.851 Iran 2.158
Arab Emirates
17 Spain 1.381 Brazil 7.606 Nigeria 1.972
18 Austria 1.289 Australia 7.566 Russia 1.838
19 Singapore 1.259 Belgium 7.471 Spain 1.670
20 Brazil 1.153 Switzerland 7.151 Pakistan 1.663
Table 3: The top 20 countries according to the average link strength
WLS ZLS LLS

WLS 1 - -
ZLS 0.796 1 -
LLS 0.599 0.716 1
Table 4: Correlations between average link strength

WLS
ZLS
LLS
WLS ZLS LLS
Figure 2: Relations between average link strength metrics
17
Page 1
4.3. Closeness Centrality
Table 4 presents the results of the top 20 countries according to the un-
weighted closeness centrality - UCC; W-weighted closeness centrality - WCC;
Z-weighted closeness centrality - ZCC; and L-weighted closeness centrality -
LCC. By means of the geodesic distance, the closeness centrality analyzes
how close the node is in relation to others. In the UCC the distance is re-
lated to the number of steps or links from a given node to reach another
node, considering the shortest path. In WCC there is a combination of the
number of links (trading partners) and the weight of the links (money flow).
The bigger the weight of the link between two nodes, the smaller is the dis-
tance between them and the WCC of both countries becomes higher.When
the importance of countries is incorporated in the network the distances are
modified.
UCC WCC ZCC LCC
(×10−3 ) (×103 ) (×109 ) (×108 )
1 United 5.650 China 3.193 China 9.557 United 1.240
States States
2 China 5.650 United 3.192 United 9.557 China 1.240
States States
3 Japan 5.650 Hong 3.189 Canada 9.554 Mexico 1.240
Kong
4 Germany 5.650 Germany 3.189 Mexico 9.554 Japan 1.240
5 United 5.650 Canada 3.188 Hong 9.551 Canada 1.240
Kingdom Kong
6 France 5.650 Mexico 3.188 Germany 9.550 United 1.240
Kingdom
7 India 5.650 Japan 3.186 Japan 9.550 Germany 1.240
8 Italy 5.650 Korea 3.186 Korea 9.548 Venezuela 1.240
9 Korea 5.650 France 3.181 United 9.544 Honduras 1.240
Kingdom
10 Australia 5.650 United 3.180 France 9.536 Iran 1.240
Kingdom
11 Spain 5.650 Netherlands 3.180 India 9.534 Indonesia 1.240
12 Netherlands 5.650 Australia 3.177 Australia 9.533 India 1.240
13 Switzerland 5.650 Italy 3.176 Italy 9.530 Australia 1.240
14 Thailand 5.650 Vietnam 3.173 Netherlands 9.528 Brazil 1.240
15 Malaysia 5.650 Poland 3.173 Brazil 9.528 Korea 1.240
16 Singapore 5.650 Austria 3.173 Vietnam 9.525 Mongolia 1.240
17 Mexico 5.618 Singapore 3.172 Belgium 9.525 France 1.240
18 Sweden 5.618 Belgium 3.171 Switzerland 9.525 Pakistan 1.240
19 Austria 5.618 Switzerland 3.171 Singapore 9.524 Italy 1.240
20 South 5.618 Malaysia 3.170 Malaysia 9.520 Turkmenistan 1.240
Africa
Table 5: The top 20 countries according to the closeness centrality
Clearly, the 16 countries most central in the UDC are the most central
also in the UCC, as they are directly linked to all other countries. The top
two positions are occupied by China and the United States. It is observed
that Hong Kong does not figure among the top 20 according to UCC or LCC
but, appears in the WCC and the ZCC with the third and the fifth positions,
respectively. This indicates that the weight of the edge and the inclusion of
18
the GDP as the node’s attributes contribute to shorten the distances among
Hong Kong and other countries in the network, i.e., the paths with greater
money flow are preferable to those of smaller flows, but certainly, for Hong
Kong, LCC did not generate results that highlighted it. Canada and Mexico,
for example, were also other countries to get a better position in closeness
centrality with the inclusion of the weights of the edges and of the nodes’
attributes, probably because their strong relation with the United States, the
country with higher GDP.
Figure 3 shows scatter plots for the closeness centrality metrics. The
relations UCC-WCC and UCC-ZCC have similar behavior. It can also be
observed that if a country has a low WCC or ZCC, it necessarily has a
low UCC. On the other hand, there are countries with high WCC or ZCC
that have low UCC. Probably such countries either have high volume links
or are linked with countries with higher GDP. In LCC the values are more
homogeneous, making it more difficult to discriminate countries. Table 6
shows the correlations between these metrics.
UCC
WCC
ZCC
LCC
UCC WCC ZCC LCC
Figure 3: Relations between the closeness centrality metrics
19 Page 1
UCC WCC ZCC LCC
UCC 1 - - -
WCC 0.565 1 - -
ZCC 0.524 0.828 1 -
LCC 0.374 0.677 0.759 1
Table 6: Correlations between the closeness centrality metrics
4.4. Betweenness Centrality

Table 7 presents the top 20 countries according to the unweighted be-
tweenness centrality - UBC; W-weighted betweenness centrality - WBC; Z-
weighted betweenness centrality - ZBC; and L-weighted betweenness central-
ity - LBC. An intermediate node acts as a facilitator between two nodes
that are not connected. When considering the weight of the edge (W, Z or
L) increases the importance of the longer paths, composed of strong links,
over shorter and weaker paths. Thus, the distance of the shortest path is
reduced. Moreover, by using real-valued weights, the chance of existing ties
in the shortest pathes reduces, thus there are fewer shortest paths, shrinking
the number of intermediate nodes. Only a few countries, especially China,
the United States and Germany, with high UBC are intermediate countries
in the WBC, ZBC and LBC, the exception being Russia and South Africa
that do not figure out in the first 20 positions according to UBC, but Rus-
sia occupies the 6-th position according to ZBC, and looses importance in
LBC, and South Africa occupies 7-th position according to LBC and looses
importance in ZBC.
China, Germany and United States figure the top three positions in the
WBC and ZBC. Germany presents a WBC greater than that of the United
States, in this way, the Germany intermediates nodes in paths whose links
have higher weights. In the ZBC, the United States surpass Germany, re-
vealing the nodes that the United States intermediate follow paths with the
most important nodes, with higher GDP. Only the first 6 countries in ZBC
and 7 in LBC are intermediates in any shortest paths. In WBC, 28 countries
have betweenness centrality greater than zero and in UBC only 8 countries
are not in the shortest path between any two countries.
Figure 4 shows scatter plots for the betweenness centrality metrics. The
weight of the edges and the insertion of the node’s attributes provide more
homogeneous results, as most countries have 0 weighted betweeness cen-
trality. Thus, the use of weighted betweeness centrality must be used with
caution, if one wants to discriminate countries. Table 8 shows the correlation
between these metrics, again the highest correlation observed was between
20
UBC WBC ZBC LBC
(×10−3 ) (×10−2 ) (×10−2 ) (×10−2 )
States States
2 China 5.821 Germany 45.82 United 72.56 China 47.42
States
3 Japan 5.821 United 27.59 Germany 13.14 India 1.695
States
4 Germany 5.821 South 8.860 India 2.253 Japan 0.565
Africa
5 United 5.821 India 6.683 Japan 1.130 Germany 0.565
Kingdom
6 France 5.821 France 5.586 Russia 1.130 Italy 0.565
7 India 5.821 Russia 5.566 Canada 0.000 South 0.565
Africa
8 Italy 5.821 Spain 5.566 Mexico 0.000 Mexico 0.000
9 Korea 5.821 Italy 4.481 Hong 0.000 Canada 0.000
Kong
10 Australia 5.821 Thailand 4.481 Korea 0.000 United 0.000
Kingdom
11 Spain 5.821 Singapore 3.371 United 0.000 Spain 0.000
Kingdom
12 Netherlands 5.821 Brazil 3.371 France 0.000 Sweden 0.000
13 Switzerland 5.821 Japan 3.361 Australia 0.000 Iran 0.000
14 Thailand 5.821 Korea 2.253 Italy 0.000 Indonesia 0.000
15 Malaysia 5.821 Australia 2.253 Spain 0.000 Australia 0.000
16 Singapore 5.821 Switzerland 2.253 Brazil 0.000 Brazil 0.000
17 New 5.718 Portugal 2.253 Vietnam 0.000 Korea 0.000
Zealand
18 Mexico 5.640 New 2.253 Belgium 0.000 Mongolia 0.000
Zealand
19 Philippines 5.595 Hong 1.130 Sweden 0.000 France 0.000
Kong
20 Sweden 5.584 Netherlands 1.130 Singapore 0.000 Pakistan 0.000
Table 7: The top 20 countries with the highest betweenness centrality
ZBC and LBC, implying that no matter what method is used to incorporated
the node’s attributes, the results show some consistency.
21
UBC
WBC
ZBC
LBC
UBC WBC ZBC LBC
Figure 4: Relations between the betweenness centrality metrics
UBC WBC ZBC LBC

UBC 1 - - -
WBC 0.479 1 - -
ZBC 0.228 0.470 1 -
LBC 0.261 0.487 0.765 1
Table 8: Correlations between the betweenness centrality metrics

Page 1
4.5. Eigenvector Centrality

Table 9 shows the top 20 countries according to the eigenvector centrality:
unweighted eigenvector centrality - UEC; W-weighted eigenvector centrality
- WEC; Z-weighted eigenvector centrality - ZEC; and L-weighted eigenvector
centrality - LEC.
By eigenvector centrality, the countries are central because they have
links with other central countries. The United States is the most central
country in this metric. As it can be seen from Table 9, when we take into
consideration the edges’ weights and also the nodes’ attributes, the variability
of the eigenvector centrality among the top 20 countries increases.
Comparing Germany and China, they figure first place according to UEC
since they are connected to all other countries of the network. However,
Germany figures the third and sixth positions according to WEC and ZEC,
respectively, while China figures the fourth and second position according
to these metrics, respectively. Thus, while taking into account the nodes’
22
attributes China surpasses Germany. We emphasize that both of these coun-
tries establish relations with all others on the network, the differences in
centrality comes from the different money flows exchanged with countries of
higher GDP. In this case, China has established larger trade relations (mone-
tary) with more central countries. It is also worth pointing out that although
Canada does not figure out in the top 20 positions according to UEC, it occu-
pies the second and third positions according to WEC and ZEC, respectively,
showing the importance of taking into account weights while analyzing eigen-
vector centrality.
UEC WEC ZEC LEC
(×10−2 ) (×10−2 ) (×10−2 ) (×10−2 )
1 United 9.750 United 59.07 United 65.43 United 70.84
2 China 9.750 Canada 46.18 China 53.81 China 62.06
3 Japan 9.750 Germany 34.29 Canada 28.24 Japan 20.08
4 Germany 9.750 China 25.02 Mexico 25.57 Germany 15.37
5 United 9.750 France 22.26 Japan 19.76 United 10.46
Kingdom Kingdom
6 France 9.750 Japan 22.25 Germany 16.05 Canada 9.180
7 India 9.750 Hong 18.98 Hong 12.86 France 7.567
Kong Kong
8 Italy 9.750 Netherlands 17.88 Korea 12.23 Mexico 6.501
9 Korea 9.750 United 10.38 United 9.511 India 6.393
Kingdom Kingdom
10 Australia 9.750 Switzerland 9.857 France 6.155 Korea 5.998
11 Spain 9.750 Italy 8.246 India 5.338 Italy 5.265
12 Netherlands 9.750 Brazil 8.000 Netherlands 5.207 Brazil 5.186
13 Switzerland 9.750 Poland 7.833 Italy 4.639 Australia 3.593
14 Thailand 9.750 Belgium 7.443 Brazil 4.629 Russia 3.345
15 Malaysia 9.750 Singapore 7.442 Australia 4.599 Spain 2.827
16 Singapore 9.750 Austria 7.286 Singapore 4.436 Netherlands 2.326
17 Ireland 9.730 Korea 6.809 Vietnam 4.279 Indonesia 2.160
18 South 9.728 Russia 6.683 Malaysia 3.732 Switzerland 1.853
Africa
19 Denmark 9.728 India 6.246 Switzerland 3.710 Saudi 1.622
Arabia
20 Sweden 9.728 Mexico 5.984 Thailand 3.582 Turkey 1.548
Table 9: The top 20 countries according to the eigenvector centrality
Figure 5 shows scatter plots among the eigenvector centrality metrics. We

can see that there is a positive relationship between the metrics. WEC is the
one with highest correlation with UEC. Regarding the relations between UEC
and ZEC and between UEC and LEC, the formation of groups is observed.
The rationale for this effect is the fact that countries with lower trade and
high volume of transactions and countries with greater trade, but with a
low volume of transactions will be seen in a equivalent manner. However, in
the UEC-LEC association, one group is isolated, distorting the relationship
between the variables. This fact also occurs in other relationships when LEC
is one of the variables. In general, the WEC, ZEC, and LEC metrics can
differentiate countries that get the same UEC value. Therefore, the inclusion
23
of more information in the network contributes positively to the distinction
of countries. Table 10 shows the correlation between these metrics, again
the highest correlation observed was between ZEC and LEC, implying that
no matter what method is used to incorporated the node’s attributes, the
results show some consistency.
UEC
WEC
ZEC
LEC
UEC WEC ZEC LEC
Figure 5: Relations between the eigenvector centrality metrics
UEC WEC ZEC LEC

UEC 1 - - -
WEC 0.716 1 - -
ZEC 0.547 0.772 1 -
LEC 0.607 0.771 0.808 1
Table 10: Correlations among the eigenvector centrality metrics

Page 1
4.6. PageRank
Table 11 shows the top 20 countries according to the unweighted PageR-
ank - UPR; W-weighted PageRank - WPR; Z-weighted PageRank - ZPR;
and L-weighted PageRank - LPR.
PageRank measures the importance of the country. A country will have
a high PageRank if it is connected to many countries or to some countries
with high PageRank. When considering the weighted network or modified
by countries’ GDP, there was no change in the importance of the first two
24
countries: United States and China. Others who did not figure the top 20
became present in the ranking when the weights were included.
It is also worth pointing out that as we consider the edges’ weights and
nodes’ attributes the variability among the PageRank values of the top 20
countries increases.
UPR WPR ZPR LPR
(×10−3 ) (×10−2 ) (×10−3 ) (×10−3 )
1 United 8.324 United 10.00 United 249.4 United 263.5
2 China 8.324 China 8.902 China 193.5 China 209.9
3 Japan 8.324 Germany 8.406 Germany 51.25 Japan 51.84
4 Germany 8.324 France 4.758 Japan 40.27 Germany 47.04
5 United 8.324 Hong 4.163 Canada 37.52 United 30.51
Kingdom Kong Kingdom
6 France 8.324 Japan 3.792 Mexico 33.80 France 24.86
7 India 8.324 Canada 3.751 United 24.15 India 20.99
Kingdom
8 Italy 8.324 United 3.446 Korea 22.42 Canada 20.48
Kingdom
9 Korea 8.324 Netherlands 3.322 Hong 21.57 Italy 17.79
Kong
10 Australia 8.324 Italy 2.934 France 20.45 Korea 16.43
11 Spain 8.324 India 2.469 Italy 14.87 Mexico 14.74
12 Netherlands 8.324 Switzerland 2.243 India 14.10 Brazil 14.59
13 Switzerland 8.324 Korea 2.068 Netherlands 13.38 Russia 11.13
14 Thailand 8.324 Belgium 1.930 Brazil 9.214 Australia 10.26
15 Malaysia 8.324 Russia 1.814 Australia 9.143 Spain 10.09
16 Singapore 8.324 Spain 1.695 Belgium 8.802 Netherlands 8.068
17 New 8.279 Poland 1.672 Russia 8.780 Indonesia 6.430
Zealand
18 Mexico 8.275 Singapore 1.529 Switzerland 8.555 Switzerland 6.308
19 Sweden 8.273 Austria 1.358 Singapore 8.359 Turkey 5.751
20 Austria 8.273 Brazil 1.322 Spain 8.065 Saudi 5.257
Arabia
Table 11: The top 20 countries according to the PageRank
Figure 6 shows scatter plots for the PageRank metrics. The different rela-
tionships between these metrics present common characteristics. WPR, ZPR
and LPR have a saturation level for low values while UPR has a saturation
level for high values. Table 12 shows the correlations between these metrics,
again the highest correlation observed was between ZPR and LPR, implying
that no matter what method is used to incorporated the node’s attributes,
the results show some consistency.
25
UPR
WPR
ZPR
LPR
UPR WPR ZPR LPR
Figure 6: Relations between the PageRank metrics
UPR WPR ZPR LPR

UPR 1 - - -
WPR 0.761 1 - -
ZPR 0.578 0.754 1 -
LPR 0.619 0.764 0.839 1
Table 12: Correlations between the PageRank

Page 1
4.7. Local Clustering Coefficient

Table 13 presents the top 20 countries according to the local unweighted
clustering coefficient - ULC; W-weighted local clustering coefficient - WLC
and; Z-weighted local clustering coefficient - ZLC. Since for directed net-
works, there are several definitions for local clustering coefficient which differ
according to the directions of the edges between adjacent nodes, we do not
explore the L-weighted version of the local clustering coefficient.
The local clustering coefficient of a country is the ratio of the number of
triangles that contains that country and the number of possible edges between
the countries adjacent to it. The number of triangles measures how many
times the country’s trading partners negotiate with each other. The weighted
local clustering coefficient measures the strength of trade flows within that
triangle.
26
The first 20 positions in the ULC contain countries that did not appeared
in the previous tables. These are countries with few trade partnerships and
low money flow. However, they have high values of ULC, indicating that
their trading partners also have trading relations to each other, i.e., they
have an active neighborhood. The first 16 countries listed in the rankings of
the previous unweighted metrics are the 16 last countries in the ranking of the
ULC. The top 20 countries according to the WLC and or the ZLC already
appeared in the previous tables and are countries with a great number of
links or big money flow.
ULC WLC ZLC
Rank Country Value Country Value Country Value
(×10−1 ) (×10−3 ) (×10−4 )
1 Antigua 10.00 Germany 11.74 United 13.23
and Barbuda States
2 Maldives 10.00 United 11.68 China 11.67
States
3 Vanuatu 10.00 China 11.45 Germany 5.108
4 Tonga 10.00 France 8.782 Japan 3.890
5 Micronesia 10.00 United 6.988 United 3.089
Kingdom Kingdom
6 Timor-Leste 10.00 Italy 6.288 Vietnam 3.047
7 Samoa 10.00 Japan 6.155 United 3.036
Arab Emirates
8 Palau 10.00 Netherlands 6.121 Indonesia 2.992
9 Kiribati 9.991 India 4.976 France 2.694
10 Tajikistan 9.990 Korea 4.552 Korea 2.338
11 Brunei 9.984 Hong 4.332 Italy 2.224
Darussalam Kong
12 Marshall 9.984 United 4.329 India 1.990
Islands Arab Emirates
13 Solomon Islands 9.973 Belgium 4.311 Canada 1.905
14 Grenada 9.972 Switzerland 4.291 Netherlands 1.876
15 Papua 9.972 Russia 4.249 Mexico 1.585
New Guinea
16 Guinea-Bissau 9.972 Vietnam 4.060 Russia 1.446
17 Iraq 9.968 Spain 3.819 Hong 1.427
Kong
18 Nepal 9.968 Poland 3.788 Spain 1.417
19 Bhutan 9.967 Canada 3.715 Brazil 1.414
20 Guyana 9.966 Turkey 3.441 Belgium 1.347
Table 13: The top 20 countries according to the local clustering coefficient
Figure 7 shows scatter plots for the local clustering coefficient metrics. It
is observed a tendency that WLC and ZLC decrease when ULC increases.
However, many countries with a great amplitude of WLC and ZLC values
have approximately the same ULC value. In the graphs it is observed that
there is no country with ULC equal to zero. Thus, every country has at
least two trading partners who also have transactions with each other and,
moreover, since the minimum ULC is above 0.6, for every country, it is more
likely that two trading partners also trade with each other than not. Table
14 presents the correlations between these metrics.
27
ULC
WLC
ZLC
ULC WLC ZLC
Figure 7: Relation between the local clustering coefficient metrics
ULC WLC ZLC

ULC 1 - -
WLC -0.569 1 -
ZLC -0.369 0.764 1
Table 14: Correlations between the local clustering coefficient metric
Page 1
4.8. Utility
The utility or benefit of a country of belonging to the network structure
studied is presented in Table 15, for the top 20 countries. The utility was
obtained in four distinct ways: unweighted utility - UU; W-weighted utility
- WU; Z-weighted utility - ZU; and L-weighted utility - LU.
As they are directly connected to all countries in the network, the top 16
countries according to the UU have the same utility. China, United States
and Germany remain in the same positions on WU, ZU and LU.
28
UU WU ZU LU
States States
2 China 2.830 United 20.29 United 58.58 China 58.27
States States
3 Japan 2.830 Germany 11.02 Germany 12.42 Germany 9.386
4 Germany 2.830 India 10.01 Japan 7.128 Japan 8.552
5 United 2.830 France 7.631 India 6.196 India 7.155
Kingdom
6 France 2.830 Japan 7.207 France 5.722 France 6.049
7 India 2.830 Italy 6.911 Italy 4.840 United 5.632
Kingdom
8 Italy 2.830 Korea 6.241 United 4.809 Italy 4.737
Kingdom
9 Korea 2.830 Russia 5.909 Russia 3.475 Spain 3.434
10 Australia 2.830 South 5.877 Korea 2.921 Netherlands 3.140
Africa
11 Spain 2.830 Netherlands 5.754 Spain 2.614 Korea 3.002
12 Netherlands 2.830 United 5.473 Netherlands 2.362 Thailand 2.882
Kingdom
13 Switzerland 2.830 Singapore 5.158 Brazil 2.186 Finland 2.646
14 Thailand 2.830 Belgium 4.906 South 1.913 Australia 2.618
Africa
15 Malaysia 2.830 Spain 4.890 Canada 1.903 Switzerland 2.281
16 Singapore 2.830 Thailand 4.848 Turkey 1.851 Mexico 2.263
17 New 2.816 Turkey 4.308 Belgium 1.783 Sweden 2.003
Zealand
18 Mexico 2.809 Switzerland 3.807 Australia 1.740 Malaysia 1.998
19 Sweden 2.804 Brazil 3.566 Switzerland 1.697 Austria 1.850
20 Austria 2.804 Australia 3.263 Mexico 1.580 Turkey 1.702
Table 15: The top 20 countries with the highest Utilities

UU
WU
ZU
LU
UU WU ZU LU
Figure 8: Relations between the utility metrics
Figure 8 shows scatter plots between the utility metrics. Only countries
with high UU feature high WU, ZU and LU. Moreover, there exists a satu-
ration level for low WU, ZU and LU values and for high UU values. Table
29
Page 1
16 presents the correlations between these metrics, again the highest cor-
relation observed was between ZU and LU, implying that no matter what
method is used to incorporated the node’s attributes, the results show some
consistency.
UU WU ZU LU
UU 1 - - -
WU 0.734 1 - -
ZU 0.658 0.835 1 -
LU 0.668 0.808 0.892 1
Table 16: Correlations between the utility metrics
4.9. Efficiency Centrality

The efficiency centrality of a country in the network structure studied
is presented in Table 17, for the top 20 countries. The EffC was obtained
in four distinct ways: unweighted EffC - UEF; W-weighted EffC - WEF;
Z-weighted EffC - ZEF; and L-weighted EffC - LEF.
The efficiency centrality measures the contribution of each node to the
network efficiency. And it is obtained when the edges of a given node are
removed. United States, China and Germany generate the largest drops of
the efficiency of the network and therefore have the highest EffC. Interesting
fact of this metric is the emphasis given to Mexico, especially in LEF. This
occurs probably due to the fact that Mexico has the United States and China
as its main trading partners.
Figure 9 shows scatter plots between the EffC metrics. We also observed
in this metric the formation of groups of countries. The high (low) number of
edges of a node compensates for the weakness (strength) of the links or the
strength (weakness) of the links compensates for the low (high) importance
of the node. There is also a discrepant group of countries, these have the
smallest GDPs and have few connections. Table 18 presents the correlations
between these metrics, again the highest correlation observed was between
ZEF and LEF, implying that no matter what method is used to incorporated
the node’s attributes, the results show some consistency.
30
UEF WEF ZEF LEF
1 United 0.014 China 0.256 United 0.397 United 0.505
States States States
2 China 0.014 Germany 0.182 China 0.303 China 0.320
3 Japan 0.014 United 0.180 Germany 0.080 Mexico 0.073
States
4 Germany 0.014 Hong 0.054 Mexico 0.077 Japan 0.064
Kong
5 United 0.014 Mexico 0.053 Canada 0.077 Canada 0.056
Kingdom
6 France 0.014 Canada 0.053 Hong 0.060 Germany 0.055
Kong
7 India 0.014 Japan 0.047 Japan 0.058 United 0.052
Kingdom
8 Italy 0.014 Korea 0.047 Korea 0.052 India 0.047
9 Korea 0.014 Russia 0.044 United 0.047 Indonesia 0.046
Kingdom
10 Australia 0.014 France 0.043 France 0.038 Australia 0.044
11 Spain 0.014 Netherlands 0.040 India 0.036 Brazil 0.044
12 Netherlands 0.014 Italy 0.040 Australia 0.036 France 0.043
13 Switzerland 0.014 United 0.038 Italy 0.034 Korea 0.041
Kingdom
14 Thailand 0.014 South 0.036 Brazil 0.033 Venezuela 0.040
Africa
15 Malaysia 0.014 India 0.034 Netherlands 0.033 Italy 0.038
16 Singapore 0.014 Australia 0.034 Belgium 0.032 Iran 0.033
17 Mexico 0.013 Spain 0.033 Switzerland 0.032 Pakistan 0.032
18 Sweden 0.013 Austria 0.032 Vietnam 0.031 Nigeria 0.031
19 Austria 0.013 Poland 0.032 Singapore 0.031 Russia 0.029
20 South Africa 0.013 Brazil 0.032 Malaysia 0.029 Spain 0.029
Table 17: The top 20 countries with the highest Efficiency Centrality
UEF
WEF
ZEF
LEF
UEF WEF ZEF LEF
Figure 9: Relations between the efficiency centrality metrics
31
Page 1
UEF WEF ZEF LEF
UEF 1 - - -
WEF 0.565 1 - -
ZEF 0.518 0.824 1 -
LEF 0.420 0.717 0.789 1
Table 18: Correlations between the efficiency centrality metrics
4.10. Reduced set of metrics

The metrics presented in the previous section are able to highlight the
most important nodes in the network, which represent the main countries
in an international trade network. Each one of the metrics has different
characteristics, however, most of them yielded similar results regarding the
classifications of countries. The world’s largest economies are among the top
20 positions presented.
Aiming to reduce the number of variables (metrics) with the lowest pos-
sible loss of information, we used the Principal Component Analysis (PCA).
The PCA is a multivariate statistical technique, widely used for the reduc-
tion of unsupervised dimension. It consists of transforming a set of original
variables into another set of variables of the same dimension called principal
components. These principal components are a linear combination of all the
original variables, are independent of each other and are estimated with the
purpose of retaining, in order of estimation, the maximum information, in
terms of the total variation contained in the data [29].
The PCA analysis was performed in two ways, in order to verify the
effectiveness of the proposed method. First, we apply the PCA with the
standardized variables: unweighted, weighted and weighted with the inclu-
sion of the node’s attributes by our proposed method. Second, replacing our
proposed method by Liu et al.’s method.
For the first case, three components were generated. Table 19 presents
the results. It can be noticed that in the unrotated component matrix, there
are high loads of the variables in more than one component. Hair [29] argues
that the same variable cannot contribute to the construction of different
components since it causes problems in the meaning of the components. In
order to avoid this problem, a rotation of components is then required. This
is a process that adjusts the axis of the components in order to obtain a
simpler and theoretically more significant solution. We use the Varimax
rotation method, which tries to maximize the dispersion of the loads within
the components and, in this way, seeks to minimize the number of variables
that have high loads in each component.
32
In the rotated matrix, we observed with greater clarity the relation of the
variables with the components, improving interpretation. The three main
components of the rotated matrix represent approximately 92.568% of the
variation. The first component explained 52.218% of the total variation and is
mainly identified by edge-weighted metrics and edge-weighted metrics which
includes node’s attributes. The metrics ZPR, ZDC, ZLS, ZEF, ZU, ZBC and
ZEC, in this order, are highly correlated with the first component, highlight-
ing their importance in the network description. The second component,
responsible for 32.553% of the total variation, is associated with unweighted
metrics UU, UCC, UPR, UDC, UEF, UBC and UEC. The third component
responsible for the lowest percentage of the total variation, 7.797%, relates,
expressively, only to WCC and ZCC, both closeness centrality metrics.
Table 19: The first three principal components
Component Matrix Rotated Component

Matrix (Varimax)
Component Component
Variable 1 2 3 1 2 3
UDC 0.636 0.756 -0.112 0.157 0.965 0.181
WDC 0.922 -0.162 -0.077 0.874 0.343 0.022
ZDC 0.882 -0.418 0.026 0.971 0.083 0.049
UCC 0.659 0.732 -0.153 0.189 0.969 0.139
WCC 0.439 0.539 0.691 0.102 0.453 0.863
ZCC 0.413 0.509 0.729 0.095 0.404 0.889
UBC 0.693 0.647 -0.215 0.261 0.934 0.063
WBC 0.777 -0.355 -0.002 0.848 0.092 0.023
ZBC 0.799 -0.499 0.049 0.942 -0.030 0.038
UEC 0.617 0.767 -0.044 0.135 0.944 0.246
WEC 0.849 -0.157 -0.076 0.809 0.312 0.013
ZEC 0.893 -0.337 0.015 0.939 0.159 0.061
UPR 0.639 0.753 -0.119 0.160 0.966 0.175
WPR 0.933 -0.164 -0.074 0.884 0.346 0.026
ZPR 0.882 -0.425 0.027 0.975 0.078 0.048
WLS 0.922 -0.161 -0.069 0.873 0.342 0.030
ZLS 0.879 -0.422 0.040 0.971 0.075 0.061
UU 0.654 0.732 -0.176 0.184 0.973 0.117
WU 0.919 -0.230 -0.018 0.907 0.269 0.061
ZU 0.834 -0.467 0.037 0.955 0.016 0.039
UEF 0.636 0.756 -0.112 0.157 0.965 0.181
WEF 0.933 -0.202 0.060 0.905 0.275 0.144
ZEF 0.913 -0.337 0.077 0.956 0.150 0.123
% of cumulative 61.762 87.634 92.568 52.218 84.771 92.568
variance
33
2 Netherlands
France
Korea Japan
Italy
Canada Germany
1
India Hong Kong
United Kingdom
2 nd component
Mexico
0
United States
China
-1
-2
-2 0 2 4 6 8 10
1 st component
Figure 10: Graph of the scores of the first two principal components
Figure 10 shows the graph of the scores for the first two principal compo-
nents, where the coordinates of each point in the graph represent the value
of the two main components obtained by the PCA for a given country. The
scores were obtained by the Anderson-Rubin method, which ensures the or-
thogonality of the estimated components, presenting zero mean and standard
deviation one. In the graph, the countries with scores higher than 0.5 in the
first component are highlighted.
For the second case, three components were also generated. Table 20
presents the results. We also see in the unrotated component matrix high
loads of variables in more than one component. Thus, a rotation of com-
ponents is necessary for a better interpretation of the contributions of the
variables to the components. The three components generated correspond-
Page 1
ing to 90.610% of the variation. The first component, which accounts for
51.350% of the total change, is mainly identified with edge-weighted met-
rics and edge-weighted metrics that includes node’s attributes. The metrics
34
LEC, LPR, LDC, LLS, LU, LEF and LBC, in that order, are highly cor-
related with the first component. The second component, responsible for
33.560% of the total variation, is associated with unweighted metrics UCC,
UU, UPR, UDC, UEF, UEC and UBC. The third component responsible for
the lowest percentage of the total variation, 5.700%, is expressively related
only to LCC.
Figure 11 shows the graph of the scores for the first two principal com-
ponents, where the coordinates of each point in the graph represent the
value of the two main components obtained by the PCA for a given country.
The scores were obtained by the Anderson-Rubin method, which ensures the
orthogonality of the estimated components (the scores are not correlated),
presenting zero mean and standard deviation one. In the graph, the countries
with scores higher than 0.5 in the first component are highlighted.
Table 20: The first three principal components
Component Matrix Rotated Component

Matrix (Varimax)
Component Component
Variable 1 2 3 1 2 3
UDC 0.617 0.776 -0.023 0.145 0.972 0.132
WDC 0.920 -0.125 -0.180 0.855 0.380 -0.139
LDC 0.883 -0.414 0.106 0.973 0.069 0.097
UCC 0.641 0.754 -0.088 0.175 0.976 0.066
WCC 0.416 0.513 0.488 0.114 0.566 0.584
ZCC 0.178 0.339 0.791 0.002 0.249 0.843
UBC 0.676 0.672 -0.198 0.244 0.941 -0.052
WBC 0.777 -0.313 -0.062 0.827 0.130 -0.060
LBC 0.767 -0.475 0.175 0.905 -0.052 0.149
UEC 0.598 0.783 0.050 0.127 0.957 0.205
WEC 0.845 -0.124 -0.163 0.789 0.341 -0.127
LEC 0.898 -0.404 0.066 0.981 0.092 0.060
UPR 0.620 0.774 -0.031 0.149 0.973 0.124
WPR 0.931 -0.127 -0.172 0.866 0.383 -0.131
LPR 0.884 -0.426 0.089 0.981 0.062 0.079
WLS 0.919 -0.125 -0.176 0.854 0.380 -0.135
LLS 0.875 -0.423 0.123 0.972 0.055 0.113
UU 0.636 0.754 -0.098 0.170 0.975 0.056
WU 0.926 -0.201 -0.050 0.901 0.297 -0.022
LU 0.862 -0.442 0.101 0.970 0.036 0.087
UEF 0.617 0.776 -0.023 0.145 0.972 0.132
WEF 0.927 -0.166 -0.052 0.885 0.328 -0.018
LEF 0.874 -0.389 0.154 0.954 0.078 0.148
% of cumulative 60.023 85.580 90.610 51.350 84.910 90,610
variance
35
2 Hong Kong
Italy France
Germany
Netherlands
Canada
Japan
1 India United Kingdom
2 nd component
0
China
United States
-1
-2
-2 0 2 4 6 8 10
1 st component
Figure 11: Graph of the scores of the first two principal components
The two analysis obtained very similar results, implying that the chosen
method to incorporate node’s attributes does not change significantly the
results. In both cases, the metrics including the node’s attributes were highly
correlated with the first component. In both analysis, USA, China and to
a lesser extent, Germany, are highlighted as the most powerful countries in
the international trade network. Thus, our proposed method has all the
advantages of Liu et al.’s method and does not suffer from the drawbacks of
Liu et al.’s method, as pointed out in Section 3.
5. Conclusions
This work explored with a greater extent the complexities of Page
a social
1
network. Motivated by the gap in the literature, where the approach to so-
cial network analysis is focused in binary networks and concentrated only in
the topological structure, this work proposed another method to introduce
36
information about the nodes in the network analysis. Such information char-
acterizes the importance of the nodes in the form of attributes. Thereby, the
study analyzed the international trade network modified by incorporating
in its structure the node’s attributes defined by the country’s GDP. Thus,
the network analysis has been approached in four ways: without considering
any weights, that is, considering only the presence or absence of trade rela-
tions; weighted, considering only the edges weights, that, in this network, is
the money flow between trading partners and; two modified weighted net-
work, where the node’s attributes is inserted in the edges weights through
our proposed method and through Liu et al.’s method.
The SNA metrics used were: the degree centrality, the average link
strength, the closeness centrality, betweenness centrality, the eigenvector
centrality, PageRank, the local clustering coefficient, utility and efficiency
centrality. Tables containing the results of the top 20 countries according to
these metrics were presented. An analysis of the positional changes of the
countries according to the four types of networks (unweighted, edge-weighted
and two modified by node’s attributes edge-weighted) was made. The met-
rics of the modified weighted network, with the inclusion of the attributes
of the nodes, were obtained by our proposed method and by the Liu et al.’s
method. In the latter, the network becomes directed. As a byproduct of
this work, we propose a definition of the utility metric for directed networks
and propose a definition of efficiency centrality for weighted and directed
networks.
Scatter plots between the calculated metrics in different ways were also
presented. These graphs facilitated and completed the understanding of the
positional changes of the countries caused by the incorporation of the edges’
weights and of the nodes’ attributes in the metrics calculations. We also
presented tables showing the correlation between the metrics calculated in
those four different ways. Most of the correlations were high and correla-
tions between unweighted metrics and metrics weighted only with the edges
weights were always higher than the correlations between unweighted metrics
and metrics weighted with both edges and node’s attributes. Therefore, the
incorporation of the nodes’ attributes does add information regarding the
countries’ SNA metrics rankings.
The incorporation of the importance of the nodes in the network struc-
ture is a way to add more information to the network. Thus the influence
or the position of the countries in the network, as defined by the metrics,
is combined with the importance of the country itself. But in some met-
37
rics like closeness centrality, betweenness centrality and utility, due to the
great financial flow and importance of a few countries which have numbers
extremely superior to the others, it is not possible, mainly in Liu et al.’s
method, to differentiate significantly the countries that do not present these
characteristics. Possibly in networks where there is no significant difference
among the importance of the nodes, among the force of the edges or in both,
this distortion does not occur.
After verifying the effect of the different metrics in the ranking of the main
countries in this international trade network, we used a principal component
analysis to reduce the number of variables and to verify the effects of these
linearly combined variables on components. The importance of including
weights in the metrics was also highlighted by the PCA analysis, showing that
the inclusion of the weights in the metrics increases their potential to explain
data variability, specially weighted metrics that include nodes’s attributes.
The unweighted network values the number of links, for example, the
higher the number of links is, the higher is the degree centrality. The weighted
network highlights the importance of the links’ strengths. Thus, there is no
distinction between nodes with less number links which are strong and nodes
with more links which are weak, according to the weighted degree centrality.
In metrics that use the distance of the shortest paths as the closeness central-
ity and the betweenness centrality, the edges weight increased the importance
of longer paths composed of stronger links over the shorter and weaker paths.
In the network with countries weights, the strength of the links is not defined
only by the money flow, but also by the average of the countries’ GDP that
are connected by these links.
The importance of considering the weighted network forms is noticeable
when analyzing the international trade network built in this work. For exam-
ple, as 16 of the 178 countries are connected to all others, when considering
the unweighted network, these countries are identical and obtain the same
results in all unweighted metrics. When we analyze the weighted network,
the money flow between any two countries represents the weight of the edges
connecting them. And considering such weights in the calculation of metrics,
makes it possible to investigate what are the most influential countries since
they have different weighted values.
The present work besides proposing such method of incorporating the
node’s attributes in SNA, also shed some light in the understanding of the
global international trade flows. Our results showed that developed countries
are the ones that are more connected, according to the degree centrality and
38
that the least developed countries have a more connected neighborhood, as
indicated by the unweighted local clustering coefficient. As a future work,
it is possible to analyze the impact of the unweighted and weighted metrics
according to different metrics present in the Human Development Index of
the country, for example. Another study, could be to analyze the positional
evolution of countries according to the metrics studied here and the countries’
economic growth over the years. Another important work, would be to use
some multi-criteria decision making process to aggregate the results of the
many centrality metrics that have been used in this work, similar in spirit to
what has been recently done in Bian et al. [30] that applied an AHP method
in unweighted networks. It is also important to replicate the present study in
different domains of applications to get a better understanding of the impacts
caused by the insertion of nodes’ attributes in SNA.
References
[1] M. E. Newman, The structure of scientific collaboration networks, Pro-
ceedings of the National Academy of Sciences 98 (2001) 404–409.
[2] M. E. J. Newman, Scientific collaboration networks. i. network con-

struction and fundamental results, Phys. Rev. E 64 (2001) 016131.
[3] S. Wasserman, K. Faust, Social network analysis: Methods and appli-

cations, volume 8, Cambridge university press, 1994.
[4] A. Barrat, M. Barthelemy, R. Pastor-Satorras, A. Vespignani, The ar-

chitecture of complex weighted networks, Proceedings of the National
Academy of Sciences of the United States of America 101 (2004) 3747–
3752.
[5] T. Opsahl, V. Colizza, P. Panzarasa, J. J. Ramasco, Prominence and

control: the weighted rich-club effect, Physical review letters 101 (2008)
168702.
[6] M. S. Granovetter, The strength of weak ties, American Journal of

Sociology 78 (1973) 1360–1380.
[7] M. E. Newman, Coauthorship networks and patterns of scientific col-

laboration, Proceedings of the national academy of sciences 101 (2004)
5200–5205.
39
[8] N. MEJ, Networks: an introduction, 2010.
[9] S. Schiavo, J. Reyes, G. Fagiolo, International trade and financial inte-

gration: a weighted network analysis, Quantitative Finance 10 (2010)
389–399.
[10] R. M. Marteleto, Análise de redes sociais: aplicação nos estudos de

transferência da informação, Ciência da Informação 30 (2001) 71–81.
[11] J. Liu, Y. Li, Z. Ruan, G. Fu, X. Chen, R. Sadiq, Y. Deng, A new method
to construct co-author networks, Physica A: Statistical Mechanics and
its Applications 419 (2015) 29–39.
[12] U. Brandes, A faster algorithm for betweenness centrality, Journal of

mathematical sociology 25 (2001) 163–177.
[13] L. C. Freeman, Centrality in social networks conceptual clarification,

Social Networks 1 (1978) 215–239.
[14] D. Krackhardt, The strength of strong ties: The importance of philos

in organizations, In Networks and Organizations: Structure, Form, and
Action (1992) 216–239.
[15] P. Bonacich, Power and centrality: A family of measures, American

Journal of Sociology 92 (1987) 1170–1182.
[16] M. Jackson, Social and economic networks, princeton, nj: Princeton

univ, 2008.
[17] D. J. Watts, S. H. Strogatz, Collective dynamics of small-worldnetworks,

nature 393 (1998) 440–442.
[18] J.-P. Onnela, J. Saramäki, J. Kertész, K. Kaski, Intensity and coherence

of motifs in weighted complex networks, Physical Review E 71 (2005)
065103.
[19] L. Page, S. Brin, R. Motwani, T. Winograd, The PageRank citation

ranking: Bringing order to the web., Technical Report, Stanford Info-
Lab, 1999.
[20] S. Brin, L. Page, The anatomy of a large-scale hypertextual web search

engine, Computer networks and ISDN systems 30 (1998) 107–117.
40
[21] A. M. Santos, Aplicações de Modelos de Grafos na Análise de Conflitos
e de Redes Sociais, Ph.D. thesis, Universidade Federal de Pernambuco,
2014.
[22] M. O. Jackson, A. Wolinsky, A strategic model of social and economic

networks, Journal of Economic Theory 71 (1996) 44–74.
[23] S. Wang, Y. Du, Y. Deng, A new measure of identifying influential

nodes: Efficiency centrality, Communications in Nonlinear Science and
Numerical Simulation 47 (2017) 151–163.
[24] V. Latora, M. Marchiori, Efficient behavior of small-world networks,

Physical review letters 87 (2001) 198701.
[25] E. Lazega, M.-T. Jourda, L. Mounier, R. Stofer, Catching up with big

fish in the big pond? multi-level network analysis through linked design,
Social Networks 30 (2008) 159–176.
[26] G. Fagiolo, Directed or undirected? A new index to check for direc-

tionality of relations in socio-economic networks, 2006. ArXiv preprint
physics0612017.
[27] D. P. Kennedy, G. L. Jackson, H. D. Green, T. N. Bradbury, B. R.

Karney, The analysis of duocentric social networks: A primer, Journal
of Marriage and Family 77 (2015) 295–311.
[28] D. P. Leyden, A. N. Link, D. S. Siegel, A theoretical analysis of the

role of social networks in entrepreneurship, Research Policy 43 (2014)
1157–1163.
[29] I. T. Hair, Principal component analysis and factor analysis, Principal

component analysis (2002) 150–166.
[30] T. Bian, J. Hu, Y. Deng, Identifying influential nodes in complex net-

works based on ahp, Physica A: Statistical Mechanics and its Applica-
tions 479 (2017) 422–436.
41
Highlights (for review)
• We discuss that non-weighted and weighted network analysis provide

different information;
• We propose to aggregate the nodes' attributes in the social network

analysis metrics as a means to gain further insights about a network;
• The proposed method to incorporate the nodes' weights in the social

network metrics is simple and can be applied in any kind of network;
• We analyzed an international trade network considering as nodes'

attributes the value of each country GDP;
• The use of nodes' attributes in the analysis of an international trade

network gives further insight on which are the most influential
countries in the global economy.

Attributes in International Trade

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Attributes in International Trade

Uploaded by

Copyright:

Available Formats

Accepted Manuscript

The use of nodes attributes in social network analysis with an application

Ricardo Lopes de Andrade, Leandro Chaves Rêgo

To appear in: Physica A

Received date : 8 November 2016

Preprint submitted to Physica A September 14, 2017

If the network is weighted, then the weighted degree centrality of vertex

where g w (vj , vk ) is the number of weighted shortest paths between vertex

Then, Wang et al. [23] used a measure to quantify the importance of a

∆EF EF [G] − EF [G0k ]

D(vi , vj ) = 1 − log(w(vi , vj )/Cdw (vi )), (20)

4.1. Degree Centrality

Table 1: The top 20 countries according to the degree centrality

UDC WDC ZDC LDC

Figure 1: Relations between the degree centrality metrics

UDC WDC ZDC LDC

Table 2: Correlations between the degree centrality metrics

4.2. Average Link Strength

Table 3: The top 20 countries according to the average link strength

WLS ZLS LLS

Table 4: Correlations between average link strength

WLS ZLS LLS

Figure 2: Relations between average link strength metrics

Table 5: The top 20 countries according to the closeness centrality

UCC WCC ZCC LCC

Figure 3: Relations between the closeness centrality metrics

Table 6: Correlations between the closeness centrality metrics

4.4. Betweenness Centrality

Table 7: The top 20 countries with the highest betweenness centrality

UBC WBC ZBC LBC

Figure 4: Relations between the betweenness centrality metrics

UBC WBC ZBC LBC

Table 8: Correlations between the betweenness centrality metrics

4.5. Eigenvector Centrality

Table 9: The top 20 countries according to the eigenvector centrality

Figure 5 shows scatter plots among the eigenvector centrality metrics. We

UEC WEC ZEC LEC

Figure 5: Relations between the eigenvector centrality metrics

UEC WEC ZEC LEC

Table 10: Correlations among the eigenvector centrality metrics

Table 11: The top 20 countries according to the PageRank

UPR WPR ZPR LPR

Figure 6: Relations between the PageRank metrics

UPR WPR ZPR LPR

Table 12: Correlations between the PageRank

4.7. Local Clustering Coefficient

ULC WLC ZLC

Figure 7: Relation between the local clustering coefficient metrics

ULC WLC ZLC

Table 14: Correlations between the local clustering coefficient metric

Table 15: The top 20 countries with the highest Utilities

Figure 8: Relations between the utility metrics

Table 16: Correlations between the utility metrics

4.9. Efficiency Centrality

UEF WEF ZEF LEF

Figure 9: Relations between the efficiency centrality metrics

Table 18: Correlations between the efficiency centrality metrics

4.10. Reduced set of metrics

Table 19: The first three principal components

Component Matrix Rotated Component

Table 20: The first three principal components

Component Matrix Rotated Component