Evaluating the Complementarity of Communication Tools
for Learning Platforms
Leonardo Carvalho
1
, Laura Assis
1
, Leonardo Lima
1
, Eduardo Bezerra
1
, Gustavo Guedes
1
,
Artur Ziviani
2
, Fabio Porto
2
, Rafael Barbastefano
1
and Eduardo Ogasawara
1
1
Federal Center for Technological Education of Rio de Janeiro, CEFET/RJ, Brazil
2
National Laboratory for Scientific Computing, LNCC, Brazil
Keywords:
Social Media, Evaluation, Learning Platforms, Communication Tools.
Abstract:
Due to the constant innovations in communications tools, several educational institutions are continually eva-
luating the adoption of new communication tools (NCT) for their adopted learning platforms (LP). Notably,
many educational institutions are interested in checking if NCT is bringing benefits in their teaching and le-
arning process. We can state an important problem that tackles this interest as for how to identify when NCT
is providing a significantly different complementary communication flow concerning the current communi-
cation tools (CCT) provided at LP. This paper presents the Mixed Graph Framework (MGF) to address the
problem of measuring the complementarity of an NCT in the scenario where some CCT is already established.
Since we are interested in the methodological process, we evaluated MGF using synthetic data. Our experi-
ments observed that the MGF was able to identify whether an NCT produces significant changes in the overall
communications of an LP according to some centrality measures.
1 INTRODUCTION
Communication tools are in constant evolution. They
usually change the way people collaborate with each
other. Not long ago, letters, telegrams, and other writ-
ten communications on paper were the mainstream.
However, since the beginning of the Internet, com-
munication tools were extended through e-mail. The
use of e-mail is widespread and almost ubiquitous in
educational institutions, being responsible for the ma-
jority of the communication flow inside them (Kola-
czyk, 2009).
Innovations in communication tools continue to
occur, and several new features, such as instant mes-
saging, blogs, and content management have been
developed. Recently, new opportunities to empower
communication among students-teachers have arisen
with the advent of online learning platforms (LP)
(Moore et al., 2011). There is some specialized LP,
such as Moodle that can be customized and extended
(Kellogg et al., 2014). Inside LP, instant messaging
(IM), wikis, and other social applications can be bet-
ter options for collaborative work and are, thus, gai-
ning momentum in scholarly communications tasks
(Mazzolini and Maddison, 2003).
As an emerging communication technology, on-
line social networks provide a variety of communica-
tion services such as profiles, comments, private mes-
saging, blogging, media file sharing, and instant mes-
saging. Some of these communication tools provide
their services through a mobile network (Chai and
Kim, 2012). These features are important as they help
to break existing barriers to communication among
members. They can stimulate interactions involving
student-teacher.
Due to demands of privacy and other pedagogical
decisions, educational institutions also may choose to
establish private social networks that are integrated
to LP and restricted to student-teacher of their main
courses. These tools are commonly inspired on pu-
blic social networks, such as Facebook, LinkedIn, in-
cluding Web 2.0/3.0 collaboration tools. In particu-
lar, they have to focus on educational issues, such
as improvement course interactions between teacher-
student and better integration with other educational
tools.
The choice for a particular LP and their customi-
zations can be both time-consuming and expensive.
Under this perspective and due to investments, edu-
cational institutions are concerned to measure the ef-
fective adoption of a new communication tool (NCT).
They are thus searching for an efficient way to as-
sess if an NCT is bringing benefits for their LP. We
can state the problem as of how to identify when an
NCT is providing a communication flow that is com-
plementary to the current communication tool (CCT)
142
Carvalho, L., Assis, L., Lima, L., Bezerra, E., Guedes, G., Ziviani, A., Porto, F., Barbastefano, R. and Ogasawara, E.
Evaluating the Complementarity of Communication Tools for Learning Platforms.
DOI: 10.5220/0006798701420153
In Proceedings of the 10th International Conference on Computer Supported Education (CSEDU 2018), pages 142-153
ISBN: 978-989-758-291-2
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
being used in LP.
In this paper, we address the problem of measu-
ring the complementarity of a NCT in the scenario
where some CCT is already an established in LP. In
order to do that, we present the Mixed Graph Frame-
work (MGF), which is a network obtained by a mix
of the CCT and the NCT networks, and is designed
to evaluate how a new complementar communication
tool can change the process of exchanging informa-
tion throughout the network. The proposed MGF is
based on the premise that the CCT can be conside-
red as a baseline for evaluating any other tool to im-
prove communication in educational institutions LP.
It is important to use a common standard represen-
tation of the communication flow to enable compa-
rison between them. In this work, the communica-
tion networks from the CCT and the NCT are mo-
deled as graphs, named G
c
and G
n
, respectively. From
these graphs, the MGF produces a mixed graph G
m
to
measure if a NCT is acting as a complementary tool
among students as compared with the CCT.
We have evaluated MGF using synthetic data that
represents teacher-students communication flows. In
our experiment, we assume the usage of a Moodle-
like LP that is mainly adopting course messages as
CCT and Social Network plugin as NCT. Based on
the shared messages in both tools, we compute several
metrics and conduct a statistical analysis on them to
evaluate the complementarity of the NCT. Our expe-
riments observed that the MGF was able to identify
whether an NCT produces significant changes in the
overall communication.
The remainder of the paper is organized as fol-
lows. Sections 2 and 3 present related work and the
general background, respectively. The proposed MGF
is described in Section 4. Section 5 presents our expe-
rimental evaluation. Finally, Section 6 concludes the
paper.
2 RELATED WORK
The analysis of social networks is widely explored,
and it has been studied for several years (Ngai et al.,
2015). Many of these studies focused on the informa-
tion that can be extracted from these networks ana-
lyzing their dynamics and structure. When it co-
mes to the impacts of communication tools adoption
in an educational environment, the need for study
in this area expanded in recent years (Chatti et al.,
2007; Hrastinski, 2009; Cadima et al., 2010; Kellogg
et al., 2014). These studies focus on the effects of the
usage of social network tools in LP and the learning-
teaching achievements (Siribaddana, 2014).
One of the main concerns about the adoption of
communication tools is related to the notion of being
social (Wasko et al., 2009). We can find several pu-
blications on the use of open social networks, most
of them representing the information flow as a graph.
In such background, it is possible to extract metrics
that enables data mining (Nettleton, 2013), such as
groups identification (clusters or cliques) that are re-
lated to concentrations of communication flows inside
the graph (Prado and Baranauskas, 2013). Many of
these metrics, such as cohesion and average distance,
are useful in network analysis, as they enable insights
about the communication flows (Newman, 2003).
In some studies, the authors structure and compare
social networks by analyzing the communication flow
among students of courses available in a Distance Le-
arning Scenario (Hamulic and Bijedic, 2009; Siribad-
dana, 2014). Their research showed that the data from
these social tools could be used to analyze the com-
munication flow and draw conclusions to improve the
available e-learning courses.
Some researchers have proposed frameworks for
understanding social media (Kellogg et al., 2014;
Chai and Kim, 2012) that suggest a theoretical frame-
work to understand social networking site users’ kno-
wledge contribution behavior and inter-relationships
among different research constructs adopted (Ngai
et al., 2015).
Many works analyze social networks and study its
behavior. Also, some papers propose frameworks for
these purposes. Nonetheless, as far as we know no ot-
her work suggests a Mixed Graph framework to mea-
sure if an NCT is complementary to a CCT already in
use in an LP.
3 BACKGROUND
This section presents the fundamental concepts for
the paper and is organized into three main subsecti-
ons. Section presents general graph concepts. Section
describes the major centrality-based measures that are
used as input for the performed statistical analysis.
Section presents the general statistical tests for non-
parametric data sets.
This section presents the fundamental concepts for
the paper and is organized into three main subsecti-
ons. Section 3.1 presents general graph concepts.
Section 3.2 describes the major centrality-based mea-
sures that are used as input for the performed statisti-
cal analysis. Section 3.3 presents the general statisti-
cal tests for non-parametric data sets.
Evaluating the Complementarity of Communication Tools for Learning Platforms
143
3.1 Graph Representation
Using graph theory terminology (Ahuja et al., 1993),
communication networks (such as LP and LP) can be
modeled as a weighted directed graph G(V, E), where
V is the set of |V| nodes and E is the set of |E| edges.
A node i V represents a member with a connection
point. The arcs (i, j) E, i V and j V represent a
communication link between two members.
A weight w
i j
> 0 is assigned to each edge with
ending nodes i and j and represents the amount of
communication flow between these two nodes. Since
G(V,E) is directed, it may be that w
i j
6= w
ji
. The ad-
jacency matrix A
i, j
= a
i, j
of the weighted graph G can
be defined as:
a
i j
=
w
i j
, if an edge is connecting node i to j
0, otherwise.
3.2 Graph Centrality Measures
When some problem is modeled by a graph, many
properties are associated with each node, such as dis-
tance and centrality. These properties provide a sum-
mary of the graph.
A general centrality measure is the weighted clo-
seness of a node v (Opsahl et al., 2010). If a node v re-
presents a member in an educational network, the clo-
seness of v measures how close a member is to others.
Collaborators that occupy central positions concer-
ning closeness are important in communication (Was-
serman and Faust, 1994). The weighted closeness of a
node v is computed by C
c
(v) =
1
x V \v
d(v,x)
, where
d(v,x) is the weighted geodesic distance between the
nodes v and x.
Another well-known measure is the weighted bet-
weenness centrality of a node v (Kolaczyk, 2009). It is
a measure aimed at summarizing the extent to which
a vertex is located ‘between’ other pairs of vertices.
Let us introduce some notation before formally define
the betweenness centrality. Consider arbitrary nodes
u,v V. A path P(u, v) which starts at u and finishes
at v is an ordered sequence of nodes, P(u, v) = <
u = v
1
,v
2
,.. .,v
k
= v >, such that e
i
= (v
i
,v
i+1
) E
for i = 1,..., k1. The length of the path P(u,v) is gi-
ven as the sum of the edge weights of the path and the
shortest path function s
G
(u,v) between nodes u,v V
is given by s
G
(u,v) = min
P(u,v)
k1
i=1
w
i,i+1
. The be-
tweenness centrality for any given node v V is then
given by C
b
(v) =
s6=t6=vV
σ(s,t|v)
σ(s,t)
, where σ(s,t) is the
number of paths P(s,t) of size s
G
(s,t) connecting s
and t and σ(s,t|v) is the number of shortest paths pas-
sing through vertex v.
The third class of centrality measure is the Klein-
berg centrality (Kleinberg, 1999). The main idea is
to identify nodes that correspond to hubs and autho-
rities. A hub is a node that points to many relevant
nodes, and an authority node is the one that is poin-
ted by many important nodes. Both are based on the
eigenvectors related to the largest eigenvalues of the
matrices AA
T
and A
T
A. The hub centrality of the node
v
i
, denoted here byC
h
(v
i
), is the ith entry of the vec-
tor x satisfying AA
T
x = λx, where λ is the largest
eigenvalue of AA
T
.
Similarly, the authority of a node v
i
, denoted here
by C
a
(v
i
), is the ith entry of the vector y satisfying
A
T
Ay = βy, where β is the highest eigenvalue of
A
T
A.
3.3 Statistical Analysis
The need to compare two different datasets is wide-
spread. Such comparison may vary according to the
objectives of the study. We can summarize two diffe-
rent statistical tests that are relevant to compare two
data sets: (i) distribution; and (ii) correlation (Larsen
and Marx, 2005). For each one of these scenarios,
there is a set of statistical tests that can be used. They
vary according to the distribution of the data sets.
Commonly, social medias are scale-free networks and
follow a power-law distribution. In this case, non-
parametric tests are more adequate. For the sake of
simplicity, we are going to present one statistical test
for comparing two data sets.
Mann-Whitney U test, also known as Wilcoxon
rank sum test from the difference in medians, is a
distribution analysis test. The goal of this test is to
measure the extent to which the medians of two inde-
pendent data sets are different from each other, i.e., to
check if the difference between the median of these
two data sets is significantly different from zero.
Spearman rank correlation test is a correlation
analysis test, whose goal is to test if the rank correla-
tion coefficient between two variables is significantly
different from zero. The null hypothesis establishes
zero correlation between two variables.
4 MGF
In this section, we present a framework to evaluate
the complementarity of communication tools using a
mixed graph modeling, called here Mixed Graph Fra-
mework (MGF). Algorithm 1 summarizes how the
MGF works. The first two lines (2-3) are related
to modeling graphs for communication tools and are
described in further detail in Section 4.1. Line (4)
is described in Section 4.2 and produces the mixed
CSEDU 2018 - 10th International Conference on Computer Supported Education
144
graph. Line (5) is described in Section 4.3 and com-
putes centrality measures to evaluate the complemen-
tarity of the NCT concerning the CCT.
4.1 Extract Functions
The first two activities of Algorithm 1 encompass mo-
deling graphs from the communication tools. Graphs
G
c
= (V
c
,E
c
) and G
n
= (V
n
,E
n
) are, respectively,
generated through the extraction Functions f Extract
c
and f Extract
n
that are applied over the CCT and NCT
datasets.
A node i V
c
and p V
n
corresponds to mem-
bers of their respective graphs G
c
and G
n
. An edge
e
i, j
E
c
represents a communication in CCT from
member i V
c
to member j V
c
and the edge weight
w
c
(i, j) corresponds to the number of messages ex-
changed from i to j. Similarly, an edge e
i, j
E
n
re-
presents a communication in the NCT from member i
to member j V
n
and the edge weight w
n
(i, j) corre-
sponds to the number of messages exchanged from i
to j.
Both f Extract
c
and f Extract
n
are User Defined
Functions (UDFs) that vary according to the adop-
ted communication tools. For example, if CCT corre-
sponds to course messages in an LP tool, the commu-
nication flow in the graph G
c
between two members i
and j V are measured by the number of posts messa-
ges exchanged by them, as described in Equation (1).
On the other hand, if the NCT is an LP tool, the com-
munication flow is measured by the weighted average
of comments and likes someone is interested in ex-
tracting from the LP, as described in Equation (2).
Algorithm 1: Main MGF Algorithm.
1: function MGF(D d
c
, D d
n
, e f
c
, e f
n
)
2: G
c
fExtract
c
(d
c
)
3: G
n
fExtract
n
(d
n
)
4: G
m
fMix(G
c
, G
n
)
5: return fAnalyze(G
c
, G
m
)
6: end function
1: function fAnalyze(G
c
, G
m
)
2: r
1
analyzeClosenessDist(G
c
, G
m
)
3: r
2
analyzeClosenessCorr(G
c
, G
m
)
4: r
3
analyzeBetweennessCorr(G
c
, G
m
)
5: r
4
analyzeEigenTopK(G
c
, G
m
)
6: return {r
1
,r
2
,r
3
,r
4
}
7: end function
w
c
(i, j) = |posts(i, j)| (1)
w
n
(i, j) =
β|comments(i, j)| + γ|likes(i, j)|
β + γ
(2)
Figures 1(a) and 1(b) display illustrative examples
of G
c
and G
n
, respectively. The graph G
c
is obtai-
ned by applying f Extract
c
over the D
c
dataset and
the graph G
n
is obtained by applying f Extract
n
over
D
n
dataset.
B
A E
D
C
1
2
6
(a)
B
A E
D
C
3
1 3
2
(b)
B
A E
D
C
4
3
6
2
3
(c)
ef
c
ef
n
mixgraphs
Figure 1: Communication flow: (a) G
c
extracted from the
CCT dataset; (b) G
n
extracted from NCT dataset; (c) G
m
produced by mixing G
c
with G
n
.
4.2 Mixed Graphs
Let G
m
= (V
m
,E
m
) be the mixed graph with node set
V
m
= V
c
= V
n
of order |V
m
| and edge set E
m
=
E
c
E
n
. To each edge e
i, j
E
m
a weight w
m
(i, j) is
assigned as given by Equation (3). The mixed graph
activity is described in Algorithm 2. It receives both
G
c
and G
n
as an input and builds the mixed graph
G
m
of order |V
m
| with its edges weights given by the
vector w
m
.
w
m
(i, j) = w
c
(i, j) + w
n
(i, j) (3)
Note that the graph G
m
represents the total flow of
communication provided by the two communication
tools and can be used to identify whether the NCT
is changing the communication flow or just mirro-
ring the communication flows between members in
the CCT. An example of G
m
can be observed in Fi-
gure 1(c) obtained from G
n
and G
c
.
Evaluating the Complementarity of Communication Tools for Learning Platforms
145
Algorithm 2: Mixed Graphs.
1: function fMix(G
c
, nG)
2: V
m
V
c
V
n
3: G
m
EmptyGraph(|V
m
|)
4: for i 1 to |V
m
| do
5: for j 1 to |V
m
| do
6: if i <> j then
7: w
m
(i, j) w
c
(i, j) + w
n
(i, j)
8: end if
9: end for
10: end for
11: return (mG,w
m
)
12: end function
4.3 Complementarity Analysis
The complementarity analysis computes centrality
measures of each vertex extracted from G
c
and G
m
.
These values are used to compute if such metrics from
G
c
are statistically significantly different from G
m
. In
this case, it indicates that G
n
is not simply an overlap
of G
c
, i.e., actually bringing complementarity in the
overall communication. Such an activity is described
in Algorithm 3.
Algorithm 3: Analysis of Centrality.
1: function analyzeClosenessDist(G
c
, G
m
)
2: vc
c
closeness(convertDist(G
c
))
3: vc
m
closeness(convertDist(G
m
))
4: return wilcox.test(vc
m
,vc
m
)
5: end function
1: function analyzeClosenessCorr(G
c
, G
m
)
2: vc
c
closeness(convertDist(G
c
))
3: vc
m
closeness(convertDist(G
m
))
4: return spearman.cor.test(vc
m
,vc
m
)
5: end function
1: function analyzeBetweennessCorr(G
c
, G
m
)
2: vb
c
betweenness(convertDist(G
c
))
3: vb
m
betweenness(convertDist(G
m
))
4: return spearman.cor.test(vb
c
,vb
m
)
5: end function
1: function analyzeHub(G
c
, G
m
,k)
2: ve
c
eigen(asHub(convertDist(G
c
)))
3: ve
m
eigen(asHub(convertDist(G
m
)))
4: ratio overlap(top
k
(ve
c
),top
k
(ve
m
))
5: sig hgeo(ratio,k ·|ve
c
|,(1k) ·|ve
c
|)
6: return {ratio, sig}
7: end function
It is worth mentioning that all centrality-based
measures expect a weighted adjacency matrix as an
input. However, in all built graphs (G
c
, G
n
, and G
m
),
the weight of the edges corresponds to the commu-
nication flow over a period. In this way, prior to
any centrality computation, it is important to con-
vert flows to distances since more messages, e-mails,
and post exchanges imply less distance between two
members. Such a transformation is described by
Function convertDist(w) that applies Equation (4) for
all edges in Algorithm 3.
w(i, j) =
1
w(i, j)
(4)
Functions closeness, betweenness, and Eigen, re-
spectively compute the weighted closeness, weigh-
ted betweenness, and weighted Eigen vectors measu-
res (Opsahl et al., 2010) of G
c
and G
m
. The first line
in all functions described in Algorithm 3 is to convert
the communication-based graph into a distance-based
graph according to Equation (4).
Function analyzeClosenessDist analyzes the clo-
seness centrality distribution. The goal is to compute
if the difference in the median of the closeness of each
graph is significantly different from zero. For that,
the nonparametric Wilcoxon rank sum is used (De-
vore and Berk, 2011). The intuition of this function
is to compute if the introduction of NCT changes the
amount of communication flow significantly concer-
ning the CCT.
Functions analyzeBetweennessCorr and
analyzeClosenessCorr correlate the betweenness
and the closeness centralities between G
c
and G
m
,
respectively. For that, the nonparametric Spearman
correlation test is used (Devore and Berk, 2011).
The intuition of these functions is to compute if
the introduction of the NCT changes significantly
the way people interact concerning the CCT by
analyzing the established communication flows.
This indicates if the NCT is not merely increasing
the scale of messages among persons, but if it is
changing the communication flow structure. Such
a test is complementary to analyzeClosenessDist.
We can have situations where analyzeClosenessDist
may not differ but either analyzeClosenessCorr or
analyzeBetweennessCorr may present significant
changes and vice-versa.
Function analyzeHub is also a complementary
analysis. It analyzes the influence of introducing new
edges in the communication flow. It starts by multi-
plying the adjacency matrix with its transpose targe-
ting the main hubs in the communication flow. This is
done in both graphs (G
c
and G
m
). Inside the function,
we calculate the top-k more central vertices in both
graphs and the overlap between them (same central
vertices in both graphs). We also compute the pro-
bability using the hypergeometric distribution of such
CSEDU 2018 - 10th International Conference on Computer Supported Education
146
an occurrence.
The MGF is implemented in R. Statistical tests,
such as Wilcoxon rank sum and Spearman correlation
tests, are available in many statistical packages, such
as R (Dalgaard, 2008) and were included in MGF.
5 EXPERIMENTAL EVALUATION
This section presents the evaluation of the proposed
MGF in measuring if the NCT brings complemen-
tarity to the CCT inside a Learning Platform (LP).
We used synthetic data to simulate both CCT and
NCT usage to explore the MGF under different group
configurations and educational scales. Both MGF
and experimental evaluation is made available at
https://github.com/eogasawara/mgf.
We have organized this section into three parts,
as follows. Section 5.1 discusses synthetic data pre-
paration that models LP (Newman et al., 2002). In
Section 5.2, we describe the general procedure of gro-
wth network used in the experimental evaluation. In
Section 5.3, we present a toy sample analysis to illus-
trate the benefits of MGF. In Section 5.4, we conduct
a sensitive analysis of MGF under different LP scena-
rios.
5.1 Synthetic Data Generation
Many networks can be framed in the definition of
scale-free networks (Barab
´
asi and Albert, 1999). A
network is classified as scale-free if the degree dis-
tribution of its nodes follows the power law mo-
del (Newman et al., 2002). Scale-free networks have
two general concepts: growth and preferential atta-
chment. The idea of growth points out to the con-
stant increase of the number of nodes in the network.
The preferential attachment means that the more con-
nected is a node, the more likely is that it gets new
links. The basic understanding for this second con-
cept is that a new member on the network has a hig-
her probability to interact with a person who interacts
with many people than with someone who is not so
active in the network.
The most notable feature of a scale-free network is
the existence of nodes with degree much higher than
the average degree in the network. The highest degree
nodes are often called hubs and have specific mea-
nings in each network. The presence of hubs is di-
rectly related to the robustness of the network. Most
of the nodes are not hubs, and the probability of a sig-
nificant impact on total flow with the departure of one
of these low degree nodes is very low. On the other
hand, the removal of a hub can cause a large impact on
the communication flow or even a network partition.
In the experiments presented in our work, we ge-
nerated G
c
(simulating hierarchical teacher-students
communication) and G
n
(simulating a social network
communication among all students) as scale-free net-
works. However, G
c
follows the organizational struc-
ture formed by the traditional teacher-student relati-
onship, whereas the G
n
does not impose such a con-
straint. This assumption is reasonable since most LP
are organized hierarchically (either teacher-students
or tutors-students).
Algorithm 4 generates synthetic instances of CCT
and NCT; and was implemented using poweRlaw, an
R package to create scale-free graphs. Initially, the
first three parameters k,v,e are related to generation
of the subgraphs that will form CCT graph (i.e., G
c
).
It starts by creating k subgraphs in G
c
. Each subgraph
has v nodes with e edges. After that, the most cen-
tral nodes in each subgraph, according to its closeness
centrality, are connected to each other to establish a
hierarchical communication in G
c
. In the end of G
c
build phase, this graph has |V
c
| = v
c
= v ·k nodes and
|E
c
|= e
c
= (e·k)+ k edges. Then, the NCT graph G
n
is generated with v
n
= |V
n
| nodes and e
n
= |E
n
| ed-
ges, such that v
n
= v
c
. By construction, G
n
is strictly
scale-free.
Table 1: Parameters used in the experimental evaluation.
Parameter Description
v
c
= v
n
Number of nodes in both graphs,
G
c
and G
n
k Number of groups in G
c
e
c
Number of edges (communication
flows) in G
c
e
n
Number of edges (communication
flows) in G
n
In Section 5.4, we explore three scenarios produ-
ced during synthetic data generation that correspond
to representative contexts for LP, such as the num-
ber of vertices. A small course has the number of
members greater than 10 and lower than 50, whereas
in medium course the number of members is grea-
ter than or equal to 50 and lower than 250. Additi-
onally, the number of messages and edges explored
in our study are in agreement with communications
using both online social networks (Benevenuto et al.,
2009). The scenarios adopted for LP are presented in
Table 2.
Evaluating the Complementarity of Communication Tools for Learning Platforms
147
1 FIGURES1
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
(a) G
c
(v
c
= 10,k
c
= 2,e
c
= 10)(b) G
25%
n
(v
n
= 10,e
n
= 10)(c) G
25%
m
= G
c
+ G
25%
n
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
(d) G
50%
n
(v
n
= 10,e
n
= 10)(e) G
75%
n
(v
n
= 10,e
n
= 10)(f) G
100%
n
(v
n
= 10,e
n
= 10)
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
(g) G
50%
m
= G
c
+ G
50%
n
(h) G
75%
m
= G
c
+ G
75%
n
(i) G
100%
m
= G
c
+ G
100%
n
Figure 1.
An example of current tool
G
c
(a) and new tool
G
n
(b) produced by Algorithm
??
. The mixed
graph G
m
is produced by Algorithm ?? from both G
c
and G
n
. A network growth for new tool (G
n
) with
ration equals to 25% (b), 50% (d), 75% (e), and 100% (f); with their respectively effects in producing
mixed graphs (G
m
), G
25%
m
(c), G
50%
m
(g), G
75%
m
(h), and G
100%
m
(i). The width of edges are related to their
weights
1/3
Figure 2: An example of current tool G
c
(a) and new tool G
n
(b) produced by Algorithm 4. The mixed graph G
m
is produced
by Algorithm 2 from both G
c
and G
n
. A network growth for new tool (G
n
) with ration equals to 25% (b), 50% (d), 75% (e),
and 100% (f); with their respectively effects in producing mixed graphs (G
m
), G
25%
m
(c), G
50%
m
(g), G
75%
m
(h), and G
100%
m
(i).
The width of edges are related to their weights.
5.2 Network Growth
Consider both G
c
and G
n
produced during the synt-
hetic dataset production. We can apply the MGF to
compute metrics and check if G
n
is complementary
to G
c
. However, to better explore MGF, in all expe-
rimental evaluation we analyzed G
n
using a network
growth described in Algorithm 5. The goal is to al-
low for the comprehension of the MGF behavior as
we increase G
n
from an empty graph until reaching
the entire G
n
structure. According to Algorithm 5,
the growth ratio δ filter both edge weights and the
number of edges in its entire structure according to
its weight distribution. The edge weights for w
n
are
CSEDU 2018 - 10th International Conference on Computer Supported Education
148
0
25
50
75
100
0.1
1.0
10 10 10 10 10
degree
frequence
(a)
5
10
0% 25% 50% 75% 100%
growth
degree
0
10
20
30
40
50
0% 25% 50% 75% 100%
growth
betweenness
5
10
15
20
0% 25% 50% 75% 100%
growth
closeness
(b)(c)(d)
0
25
50
75
100
0
20
40
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Gm
Gc
(e)
0e+00
1e03
2e03
2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03
Gm
Gc
( f)
Figure 2. Descriptive statistics of G
m
in the toy example grouped by growth ratio
d = {0,25,50, 75,100}. The degree distribution of G
m
is in log x log scale (a). Box-plot of degree (b),
closeness (c), and betweenness (d) distributions of G
m
. Correlation plot of betweenness (G
c
x G
m
) (e).
Correlation plot of closeness (G
c
x G
m
) (f)
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●●
●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
SE:small SE:med. SE:large 5%
●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
SE:small SE:med. SE:large 5%
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness spearman.test
config:
SE:small SE:med. SE:large 5%
(a)(b)(c)
Figure 3. Scenario of Small Enterprise - varying number of edges in G
n
: betweenness correlation
analysis (a), closeness median analysis (b), closeness correlation analysis (c)
2/3
Figure 3: Descriptive statistics of G
m
in the toy example grouped by growth ratio δ = {0,25,50,75,100}. The degree
distribution of G
m
is in log x log scale (a). Box-plot of degree (b), closeness (c), and betweenness (d) distributions of G
m
.
Correlation plot of betweenness (G
c
x G
m
) (e). Correlation plot of closeness (G
c
x G
m
) (f).
all multiplied by
δ
100
, to set the relative strength of
usage in both networks. The lesser the value of δ,
the lesser is the communication flow inside the gene-
rated NCT. Additionally, only δ percentile of edges
is presented in w
n,δ
. This allows for simulating the
increase of new relationships among members accor-
ding to time. Each combination of w
c
, w
n,δ
is used
as input for f Analyze. All metrics are collected and
stored in a result set RS. Once RS is complete, it is
possible to plot charts, such as the ones presented in
the experimental evaluation.
Note that Algorithm 4 takes as input the growth
ratio δ (0 δ 100). Initially, the edge weights for
both G
c
and G
n
are randomly generated according to
the same distribution. After that, Table 1 summarizes
parameters adopted in experimental evaluation.
5.3 Toy Sample Analysis
As an example, we present a toy graph that corre-
sponds to one of the smallest LP possible. It has ten
vertices, two groups for G
c
, and ten edges in both G
c
Evaluating the Complementarity of Communication Tools for Learning Platforms
149
0
25
50
75
100
0.1
1.0
10 10 10 10 10
degree
frequence
(a)
5
10
0% 25% 50% 75% 100%
growth
degree
0
10
20
30
40
50
0% 25% 50% 75% 100%
growth
betweenness
5
10
15
20
0% 25% 50% 75% 100%
growth
closeness
(b)(c)(d)
0
25
50
75
100
0
20
40
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Gm
Gc
(e)
0
25
50
75
100
0e+00
1e03
2e03
2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03 2.5e04 5.0e04 7.5e04 1.0e03
Gm
Gc
( f )
Figure 2. Descriptive statistics of G
m
in the toy example grouped by growth ratio
d = {0,25, 50,75,100}. The degree distribution of G
m
is in log x log scale (a). Box-plot of degree (b),
closeness (c), and betweenness (d) distributions of G
m
. Correlation plot of betweenness (G
c
x G
m
) (e).
Correlation plot of closeness (G
c
x G
m
) (f)
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
SE:small SE:med. SE:large 5%
●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
SE:small SE:med. SE:large 5%
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness spearman.test
config:
SE:small SE:med. SE:large 5%
(a)(b)(c)
Figure 3. Scenario of Small Enterprise - varying number of edges in G
n
: betweenness correlation
analysis (a), closeness median analysis (b), closeness correlation analysis (c)
2/3
Figure 4: Scenario of Small Course - varying number of edges in G
n
: betweenness correlation analysis (a), closeness median
analysis (b), closeness correlation analysis (c).
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
SE:low SE:mod. SE:high 5%
●● ●●
●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
SE:low SE:mod. SE:high 5%
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness spearman.test
config:
SE:low SE:mod. SE:high 5%
(a)(b)(c)
Figure 4. Scenario of Small Enterprise - varying number of groups in G
c
: betweenness correlation
analysis (a), closeness median analysis (b), closeness correlation analysis (c)
●● ●● ●● ●● ●● ●● ●●
●● ●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●● ●●
●● ●●
●●
●●
●● ●●
●● ●●
●●
●●
●●
●●
●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●●
●● ●● ●● ●● ●●
●●
●●
●● ●●
●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness wilcox.test
config:
ME:low ME:mod. ME:high 5%
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
ME:low ME:mod. ME:high 5%
●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
ME:low ME:mod. ME:high 5%
(a)(b)(c)
Figure 5.
Scenario of Medium Enterprise - varying both number of groups in
G
c
and number of edges in
G
n
: betweenness median analysis (a), betweenness correlation analysis (b), closeness median analysis (c)
3/3
Figure 5: Scenario of Small Course - varying number of groups in G
c
: betweenness correlation analysis (a), closeness median
analysis (b), closeness correlation analysis (c).
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
SE:low SE:mod. SE:high 5%
●● ●●
●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
SE:low SE:mod. SE:high 5%
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness spearman.test
config:
SE:low SE:mod. SE:high 5%
(a)(b)(c)
Figure 4. Scenario of Small Enterprise - varying number of groups in G
c
: betweenness correlation
analysis (a), closeness median analysis (b), closeness correlation analysis (c)
●● ●● ●● ●● ●● ●● ●●
●● ●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●● ●●
●● ●●
●●
●●
●● ●●
●● ●●
●●
●●
●●
●●
●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●●
●● ●● ●● ●● ●●
●●
●●
●● ●●
●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness wilcox.test
config:
ME:low ME:mod. ME:high 5%
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
betweenness spearman.test
config:
ME:low ME:mod. ME:high 5%
●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
●● ●●
●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
growth
closeness wilcox.test
config:
ME:low ME:mod. ME:high 5%
(a)(b)(c)
Figure 5.
Scenario of Medium Enterprise - varying both number of groups in
G
c
and number of edges in
G
n
: betweenness median analysis (a), betweenness correlation analysis (b), closeness median analysis (c)
3/3
Figure 6: Scenario of Medium Course - varying both number of groups in G
c
and number of edges in G
n
: betweenness median
analysis (a), betweenness correlation analysis (b), closeness median analysis (c).
and G
n
(v
c
= v
n
= 10, k
c
= 2, e
c
= e
n
= 10).
Figures 2(a) and 2(b) are respectively examples of the
CCT and the NCT graphs produced by Algorithm 4
according to this small setup. Figure 2(c) presents
the produced mixed graph (G
m
) from both G
c
and G
n
using Algorithm 2.
In the example, Figure 2(a) simulates communica-
tions that occur through CCT inside a small course. In
this case, we assume that the course has two groups.
It is possible to view some clusters of communication,
which can be found among students who share a close
relationship, such as work on related tasks, where the
internal processes of the course make them to have
a direct communication. Despite these clusters, it is
possible to observe that the graph is connected. This
means that with the mediation of one or more persons,
the information can be disseminated through the net-
work. In a small network like G
c
, we can visually
inspect the characteristics that are part of the goals
of our analysis, such as connectivity, the presence of
clusters, and center points connecting them which are
the students identified as 2 and 7. Clusters commu-
CSEDU 2018 - 10th International Conference on Computer Supported Education
150
Algorithm 4: Synthetic dataset production.
1: function SyntheticDatasets(k,v,e,e
n
)
2: for all i 1 to k do
3: G
i
c
new ScaleFreeGraph(v, e)
4: G
c
G
c
G
i
c
5: end for
6: for i 1 to k
E
1 do
7: for j i + 1 to k
E
do
8: e
l
connect(G
i
c
,G
j
c
)
9: E
c
E
c
e
10: end for
11: end for
12: v
c
v ·k
13: v
n
v
c
14: G
n
new ScaleFreeGraph(v
n
,e
n
)
15: return ({G
c
,G
n
})
16: end function
Table 2: LP Scenarios.
Scenario Description
SC (G
n
scale)
v
c
= 30, k
c
= 3, e
c
= 60
small : e
n
= 25
medium : e
n
= 45
large : e
n
= 55
SC (G
c
groups)
v
c
= 30, e
c
= 60, e
n
= 45
low : k
c
= 2
moderated : k
c
= 3
high : k
c
= 5
MC (G
c
groups)
v
c
= 150, e
c
= 60
low : k
c
= 10, e
n
= 120
moderated : k
c
= 15, e
n
= 180
high : k
c
= 25, e
n
= 300
nicate with each other through the central points. We
applied a similar procedure to produce the graph as-
sociated to the NCT (G
n
) depicted in Figure 2(b) and
described in Algorithm 4.
Figure 2 explores different network growth (δ) of
the new tool (G
n
) using ratios such as 25%, 50%,
75%, and 100% in Algorithm 5. It is possible to ob-
serve that both the number of edges in G
n
and their
weights are explored in different growth ratios (G
25%
n
(b), G
50%
n
(d), G
75%
n
(e), and G
100%
n
(f)). This leads to
different mixed graphs G
m
: G
25%
m
(c), G
50%
m
(g), G
75%
m
(h), and G
100%
m
(i) by mixing G
c
with G
n
. By visually
inspecting the instance of G
m
presented in Figure 2, it
seems that the hierarchical structure does not restrict
the communication flow as the growth ratio of G
n
in-
creases.
Algorithm 5: Network Growth.
1: function NetGrowth(w
c
,w
n
,r)
2: RS {}
3: for all δ 0 to 100 step r do
4: w
n,δ
Filter(δ,
δ
100
·w
n
)
5: w
m,δ
fMix(w
c
, w
n,δ
)
6: RS RS f Analyze(w
c
,w
m,δ
)
7: end for
8: plotCharts(RS)
9: end function
To better comprehend the toy sample, Figure 3
presents descriptive statistics for G
m
produced by
mixing G
c
(v
c
= 10,k
c
= 2,e
c
= 10) with G
n
(v
n
=
10,e
n
= 10). Figure 3(a) depicts the frequency of
degree of G
m
as G
n
grows. The degree of vertices
increases as G
n
grows. The plots in log x log scale
fits a power law distributions, i.e., suggesting a scale-
free graph. This behavior is also summarized in Fi-
gure 3(b). Additionally, Figures 3(c) and 3(d) des-
cribe the closeness and betweenness centrality distri-
bution. In Figure 3(d), the box plot for growth ratios
of 50%, does not present any intersection with box
plots of smaller growth ratios (0% and 25%). This
indicates significant difference among them, i.e., the
median closeness of G
50%
m
is higher than in G
c
. Ne-
vertheless, the betweenness described in Figure 3(c)
does not present any significant difference among
them.
Furthermore, Figures 3(e) and 3(f) present, re-
spectively, a scatter plot for the closeness and bet-
weenness correlation between G
c
and G
m
. The cor-
relation is plotted with a confidence interval of 95%.
It is possible to observe that both are correlated. This
indicates, for example, that although Figure 3(c) indi-
cates an increase in closeness introduced by G
n
, such
an increase does not change the topology of G
c
, i.e., it
is not introducing a complementary behavior. It is ac-
tually just introducing an increase in the scale of G
m
with respect to G
c
.
However, analyzing these plots may not be appli-
cable in general, especially for more extensive net-
works, such as in a distance learning education. To
tackle this problem, the MGF uses statistical analysis
to assess and monitor the complementarity of NCT. It
applies the Wilcoxon rank sum test and the Spearman
rank correlation test to both betweenness and close-
ness as described in our Main Analysis.
5.4 Sensitive Analysis
In this section, we evaluate the proposed MGF using
synthetic data described in Section 5.1. It is worth
Evaluating the Complementarity of Communication Tools for Learning Platforms
151
mentioning that the objective of this section is not to
assess the impacts of introducing a NCT. Instead, we
intend to evaluate whether the MGF can distinguish
G
c
and G
m
according to the influence of G
n
. We have
conducted a sensitivity analysis between networks.
The goal is to identify if the NCT keeps the communi-
cation flow provided by CCT or if it introduces alter-
native and significant changes in the communication
flows.
In the first scenario described in Table 2 we explo-
red the number of communication flows in the NCT
of courses under a small, medium, and large scale.
Regarding betweenness, Figure 4(a) indicates a signi-
ficant difference for the correlation, when the growth
ratio is greater than 60%. Additionally, in terms of
closeness, both median (Figure 4(b)) and correlation
(Figure 4(c)) presents a significant difference when
the growth ratio are greater than 40% and 55%, re-
spectively.
We also explored the second scenario for cour-
ses, in which we vary the number of groups inside
G
c
. Figure 5(a) indicates a significant difference for
the betweenness correlation when the growth increa-
ses. They were reached after a growth of 65%. In
fact, the growth threshold for a significant difference
occurs later when the group size is moderate or low.
When it comes to closeness, both median analysis (Fi-
gure 5(b)) and correlation analysis (Figure 5(c)) pre-
sent a significant differences when growth is greater
than 40%. This is interesting as it indicates an incre-
ase in the number of messages in G
m
and a difference
in the network communication topology, as well.
In our third evaluation scenario, we explored the
number of communication flows in the NCT and the
number of groups inside G
c
of a Medium Course un-
der small, medium, and large scale. Regarding be-
tweenness, as depicted in Figures 6(a) and 6(b), we
observe a significant difference for the median and
correlation as the growth ratio reaches 35% and 60%,
respectively. A similar behavior occurs with close-
ness. Figure 6(c) indicates a significant difference for
the closeness median when the growth is greater than
35%. In fact, for the medium size case, only when re-
aching an increase higher than 70% we found a clear
significant difference. Before this value, we observed
an oscillatory behavior around the significance thres-
hold.
6 CONCLUSION
This paper proposes a Mixed Graph Framework
(MGF), which aims at providing a set of quantita-
tive approaches to analyze the complementary of a
new communication tool (NCT) with relation to a cur-
rent communication tool (CCT) in a learning platform
(LP). This is done by measuring when the NCT brings
significant differences in the overall educational com-
munication flow concerning the usage of the CCT.
We model CCT and NCT communication interacti-
ons as the weighted graphs G
c
and G
n
, respectively.
From these graphs, the MGF computes a mixed graph
(G
m
) that combines both G
c
and G
n
considering their
usage. Our approach is then able to identify changes
in overall communication within the educational.
We also evaluated the proposed MGF using synt-
hetic data from which we have conducted a sensi-
tive analysis. The sensitivity analysis is used to com-
pare the weighted closeness and betweenness of both
G
c
and G
m
. Our approach can identify whether G
n
is providing any changes in the entire communica-
tion flow. It is worth mentioning that our method
does not propose adopting the NCT as a replacement
for the CCT to promote communication empower-
ment. Instead, the goal of MGF is to aid managers in
a decision-making process, giving them elements to
conduct what-if analysis while deploying NCTs and
measuring its influence in the entire set of communi-
cation solutions adopted in the educational.
We considered the evolution of a single network
over time including time notion in the proposed fra-
mework as a promising future research. As well as
performing case studies with networks of different si-
zes, which is a useful analysis for educational institu-
tions with scenarios of reorganizations, mergers and
divisions of courses.
ACKNOWLEDGMENTS
The authors would like to thank CNPq, CAPES, and
FAPERJ for partially funding this research.
REFERENCES
Ahuja, R. K., Magnanti, T. L., and Orlin, J. B. (1993). Net-
work Flows: Theory, Algorithms, and Applications.
Pearson, Englewood Cliffs, N.J, 1 edition edition.
Barab
´
asi, A.-L. and Albert, R. (1999). Emergence of sca-
ling in random networks. Science, 286(5439):509–
512.
Benevenuto, F., Rodrigues, T., Cha, M., and Almeida, V.
(2009). Characterizing user behavior in online social
networks. In Proceedings of the ACM SIGCOMM In-
ternet Measurement Conference, IMC, pages 49–62.
Cadima, R., Ferreira, C., Monguet, J., Ojeda, J., and Fer-
nandez, J. (2010). Promoting social network aware-
CSEDU 2018 - 10th International Conference on Computer Supported Education
152
ness: A social network monitoring system. Computers
and Education, 54(4):1233–1240.
Chai, S. and Kim, M. (2012). A socio-technical approach to
knowledge contribution behavior: An empirical inves-
tigation of social networking sites users. International
Journal of Information Management, 32(2):118–126.
Chatti, M., Jarke, M., and Frosch-Wilke, D. (2007). The
future of e-learning: a shift to knowledge networking
and social software. International Journal of Know-
ledge and Learning, 3(4-5):404–420.
Dalgaard, P. (2008). Introductory Statistics with R. Sprin-
ger, New York, 2nd edition edition.
Devore, J. L. and Berk, K. N. (2011). Modern Mathemati-
cal Statistics with Applications. Springer, New York ;
London, 2nd ed. 2012 edition edition.
Hamulic, I. and Bijedic, N. (2009). Social network analy-
sis in virtual learning community at faculty of infor-
mation technologies (fit), Mostar. In Procedia - So-
cial and Behavioral Sciences, volume 1, pages 2269–
2273.
Hrastinski, S. (2009). A theory of online learning as online
participation. Computers and Education, 52(1):78–
82.
Kellogg, S., Booth, S., and Oliver, K. (2014). A so-
cial network perspective on peer supported learning in
MOOCs for educators. International Review of Rese-
arch in Open and Distance Learning, 15(5):263–289.
Kleinberg, J. (1999). Authoritative sources in a hyperlinked
environment. Journal of the ACM, 46(5):604–632.
Kolaczyk, E. D. (2009). Statistical Analysis of Network
Data: Methods and Models. Springer, New York,
2009 edition edition.
Larsen, R. J. and Marx, M. L. (2005). An Introduction to
Mathematical Statistics and Its Applications. Prentice
Hall, Upper Saddle River, N.J, 4 edition edition.
Mazzolini, M. and Maddison, S. (2003). Sage, guide or
ghost? the effect of instructor intervention on student
participation in online discussion forums. Computers
and Education, 40(3):237–253.
Moore, J., Dickson-Deane, C., and Galyen, K. (2011). E-
Learning, online learning, and distance learning envi-
ronments: Are they the same? Internet and Higher
Education, 14(2):129–135.
Nettleton, D. (2013). Data mining of social networks repre-
sented as graphs. Computer Science Review, 7(1):1–
34.
Newman, M. (2003). The structure and function of complex
networks. SIAM Review, 45(2):167–256.
Newman, M., Watts, D., and Strogatz, S. (2002). Random
graph models of social networks. Proceedings of the
National Academy of Sciences of the United States of
America, 99(SUPPL. 1):2566–2572.
Ngai, E., Tao, S., and Moon, K. (2015). Social media re-
search: Theories, constructs, and conceptual frame-
works. International Journal of Information Manage-
ment, 35(1):33–44.
Opsahl, T., Agneessens, F., and Skvoretz, J. (2010). Node
centrality in weighted networks: Generalizing degree
and shortest paths. Social Networks, 32(3):245–251.
Prado, A. and Baranauskas, M. (2013). Addressing structu-
ral and dynamic features of scientific social networks
through the lens of Actor-Network Theory. Social
Network Analysis and Mining, 3(4):1263–1276.
Siribaddana, P. (2014). Making distance learning an ef-
fective health information systems training strategy:
A combined social network analysis and content ana-
lysis perspective. Electronic Journal of Information
Systems in Developing Countries, 61(1).
Wasko, M., Teigland, R., and Faraj, S. (2009). The provi-
sion of online public goods: Examining social struc-
ture in an electronic network of practice. Decision
Support Systems, 47(3):254–265.
Wasserman, S. and Faust, K. (1994). Social Network Analy-
sis: Methods and Applications. Cambridge University
Press, Cambridge ; New York, 1 edition edition.
Evaluating the Complementarity of Communication Tools for Learning Platforms
153