. Technology Enhanced Lea
, Vol. 8, No. 2, 2016 169
Copyright © 2016 Inderscience Enterprises Ltd.
An automatic adaptive grouping of learners in an
e-learning environment based on fuzzy grafting
and snap-drift clustering
Mohammad Sadegh Rezaei and
Gholam Ali Montazer*
Information Technology Department,
Tarbiat Modares University,
Jalal Ale Ahmad Highway,
P.O. Box 14115-179, Tehran, Iran
Abstract: Adaptive learning systems provide e-learning-based educational
services tailored to the needs, preferences and capabilities of learners. The
quality of services provided by these systems largely depends on their ability to
acquire proper description of learners regarding their personality, behaviour
and learning style and to categorise these learners accurately into homogeneous
and heterogeneous groups. The ability of an adaptive system to provide a
suitable course through suitable presentation is influenced by the accuracy of
mentioned grouping process. This paper presents a novel adaptive learning
system possessing an automatic and intelligent learner grouping capability. The
grouping approach used in this system consists of four stages, identifying the
group structures, classifying the learners into the identified groups, detecting
the expiration of groups and modifying the groups of learners. This clustering
concept is developed by modified fuzzy snap-drift method, and the process of
assigning suitable content to identified groups is implemented by a decision
tree. The proposed system is implemented on an e-learning course to evaluate
its effect on the learning quality. The evaluation of ‘academic satisfaction’ and
‘progress’ criteria shows that the proposed system has been able to make
significant improvements in e-learning environment.
Keywords: adaptive grouping; adaptive learning system; e-learning; fuzzy
grafting clustering; learning style; neural network; snap-drift; technology
Reference to this paper should be made as follows: Rezaei, M.S. and
Montazer, G.A. (2016) ‘An automatic adaptive grouping of learners in an
e-learning environment based on fuzzy grafting and snap-drift clustering’,
Int. J. Technology Enhanced Learning, Vol. 8, No. 2, pp.169–186.
Biographical notes: Mohammad Sadegh Rezaei received BS in Information
Technology Engineering from Shiraz University of Technology, Shiraz, Iran, in
2011 and an MS in Information Technology Engineering from Tarbiat Modares
University, Tehran, Iran, in 2013, where he is currently working towards PhD
degree in Tehran University. His research interest lies in Technology Enhanced
170 M.S. Rezaei and G.A. Montazer
Gholam Ali Montazer received BSc in Electrical Engineering from Kh.N.
Toosi University of Technology, Tehran, Iran, in 1991, an MSc in Electrical
Engineering from Tarbiat Modares University, Tehran, Iran, in 1994 and PhD
in Electrical Engineering from the same university, in 1998. He is an Associate
Professor in the Department of Information Technology at School of
Engineering, Tarbiat Modares University (TMU). His research interests lie in
the area of artificial intelligence, soft computing approaches such as artificial
neural network (ANN), fuzzy set theory (FST) and rough set theory (RST),
pattern recognition, e-learning and e-government.
1 Introduction and theoretical framework
The development of information technology and its diverse capabilities has led to advent
and spread of a new form of learning process called e-learning. With dramatically
improved access to information and application of computing technology to education,
e-learning is now expected to adapt the process of learning to its clients (Akbulut and
Cardak 2012; Essalmi et al., 2010; O’Donnell et al., 2015).
The process provided by an adaptive learning system to accommodate learning to
each learner consists of two phases:
1 identification of specific needs of each learner
2 selection and provision of courses and curriculums based on the identified needs
(Botsios, Georgiou and Safouris, 2008; O’Donnell et al., 2015).
But difficulties in precise recognition of each learner’s specific educational needs, certain
limitations on accommodation of diverse educational content to those needs and lack of
access to technologies required to support one-to-one education complicates the provision
of a learner-centred learning process, hence customised learning acts as the best
alternative to provide the most suitable service for learners with similar needs and
requirements (Jin et al., 2006; Bachari, Abelwahed and Adnani, 2011). The aim of this
study is thus to develop a new adaptive system with intelligent grouping capabilities
enabling it to automatically adapt the educational content to the specifications of learners
with similar needs and learning processes.
The rest of this paper is organised as follows: Section 2 describes the concept of
adaptive learning and its methods and reviews current adaptive e-learning systems;
Section 3 introduces the proposed adaptive learning system and its intelligent grouping
method; Section 4 describes and evaluates the process of implementing the proposed
system on an e-learning course; and the final section presents the results and conclusion.
2 Adaptive learning system
Adaptive learning refers to the process of accommodating the educational course and
curriculum of an e-learning environment to the needs, preferences and capabilities of
learners (Akbulut and Cardak, 2012; O’Donnell et al., 2015; Dominic, Xavier and
An e-learning environment based on fuzzy grafting 171
Francis, 2015; Bachari, Abelwahed and Adnani, 2011). This adaptation can be realised in
three forms: adaptive presentation of educational content, adaptive sequencing of its
components and adaptive tools of navigation (Papanikolaou et al., 2003; Dominic, Xavier
and Francis, 2015).
a Adaptive content presentation: In this approach, educational contents will be
presented according to learners’ characteristics. The aim of this process is to adapt
the educational content (in terms of type of medium) to preferences, needs and
capabilities of learners and subsequently increase the speed and quality of learning.
The Arthur (Gilbert and Han, 2002) and the CS 383 (Carver Jr, Howard and Lane,
1999; Sonwalkar, 2015) are the major adaptive systems of this type.
b Adaptive content sequencing: In this approach, the order and sequence by which
different sections of educational content will be presented will be tailored to
learners’ characteristics. Here, the terms order and sequence do not refer to rational
sequence (as its presence is a basic prerequisite to learning)but to the order of
different content components that have no prerequisite relationships with each other
and whose sequence of presentation has significant effect on learning. The ACE
(Specht and Oppermann, 1998) and the INSPIRE (Papanikolaou et al., 2003;
Dominic, Xavier and Francis, 2015; Chen, Chiu and Huang, 2015) systems are the
most important adaptive systems of this kind.
c Adaptive navigation tools: The purpose of this form of adaptation is to provide
proper orientation for the users of an e-learning environment enabling them to access
system resources based on their habits, needs and requirements. This form of
adaptation can be achieved by adjusting the visible links that determine the
orientation of learning. The AES-CS (Triantafillou, Pomportsis and Demetriadis,
2003) system provides this form of adaptation.
The subject of customised learning and adaptation of educational content in its various
forms was first introduced in 1990, but it has undergone significant improvement in
recent decade. Next, we review some of the works carried out on this topic.
The ‘learn fit’ system adapts the educational content to learners. In this system, the
Myers-Brigs learning style model is used to describe the learner, who is modelled by a
non-automatic procedure implemented through questionnaire. This system uses a
Bayesian network to classify learners in different categories and adapt the content and
curriculum to their specifications (Bachari, Abelwahed and Adnani, 2011).
The ‘i-learn’ system is another form of intelligent e-learning systems that facilitates
the task of adapting the content and curriculum to the learners. In this system, the VARK
learning style is used to model the learner, and the rule-based classification method is
used to categorise the learners and adjust the content (Peter, Bacon and Dastbaz, 2010).
In the adaptive system proposed by Minto, learners are modelled by characterisation
models (such as Catele-16-PF) and are then classified through K-means clustering
(Minetou and SYC, 2005). The adaptive system of Zheng uses a new matrix-based
clustering method to classify learners. He believes that the proposed method overcomes
the most important shortcoming of K-means, i.e., the indetermination of optimum
number of clusters. This study also uses the Catele-16-PF method to model the learners
(Zhang et al., 2007). In the system proposed by Yang et al. learners are modelled by
character models and a self-organising neural network is used to classify the learners
based on their preferences and capabilities (Zakrzewska, 2009).
172 M.S. Rezaei and G.A. Montazer
A researcher who seeks to develop a learner classification method always strive for
the best possible combination of speed and accuracy. Many researchers have combined
the conventional clustering methods such as K-means and fuzzy C-means with
optimisation methods to achieve better accuracy, but this approach leads to sharply
increased time complexity, which can have major negative impacts on the quality of the
resulting e-learning system. On the other hand, previous researches have neglected to
evaluate their clustering results by the criteria that depend on learning styles and models,
and this issue obstructs any detailed assessment on the accuracy of these learner
3 The proposed learning adaptive system
The proposed adaptive system consists of three subsystems: ‘learning management’,
‘learner grouping’ and ‘content assignment’. The architecture of this system is shown in
Figure 1 Adaptive e-learning system architecture (see online version for colours)
The ‘learning management’ subsystem is tasked with the management of educational
content and adjustment of their sequence. The ‘learner grouping’ subsystem is
responsible for automatic clustering of learners into homogeneous groups and facilitating
the provision of group-tailored educational content. The ‘content assignment’ subsystem
is tasked with establishing the rules of content adaptation for each group and customising
the learning process based on these rules. The detailed descriptions of all
abovementioned subsystems are presented in the following.
An e-learning environment based on fuzzy grafting 173
3.1 Learning management subsystem
This subsystem is tasked with maintaining educational content and presenting them to
learners. Other tasks of this subsystem include managing user access, logging user
activity, monitoring and managing registration, managing finances, setting up courses,
deriving lessons from educational content and aggregating these lessons into courses. The
learning management subsystem designed in this paper is based on the Persian version of
Moodle learning management system (v. 2.5).
3.2 Learner grouping subsystem
This subsystem is tasked with automatic modelling of learners based on the
Felder-Silverman learning style and classifying them into homogeneous groups to initiate
the adaptive learning process. The learner grouping process of this subsystem includes
four major steps:
− Initial identification of group structures: In this step, subsystem identifies the
groups of participating e-learners based on a set of clustering criteria. When learner
population is large, subsystem can use random sampling to select the learners to be
used for cluster identification.
− Assigning learners to identified groups: The main objective of this step is to assign
other learners to the classes identified in the previous step to generalise the grouping
structure to the entire population. Other objectives of this step include assigning
those learners who enter the system after structure identification and correcting the
class of those learners whose model characteristics have changed or expired. The
classifier used in this step first undergoes an automatic training phase, which enables
it to classify other learners into identified classes.
− Detecting changes in group structure: When characteristics of a classified learner
get sufficiently divergent from those of other learners (in the same class), subsystem
detects and announces an expired class structure and triggers the reidentification
procedure. In this study, subsystem uses Davies-Bouldin index to detect class
expiration. Here, the greatest value of Davies-Bouldin index at the grouping
made by basic grafting clustering methods is considered as the threshold of
− Reidentification of changed group structures: When an expiration event is
announced, subsystem uses a clustering method to identify the new structure of
The proposed grouping subsystem is composed of two processes, ‘group identification’
and ‘learner-class allocation’; one agent, ‘learner model’ and two databases, ‘learner
model’ and ‘learner group’. Figure 2 shows the architecture of this grouping subsystem.
174 M.S. Rezaei and G.A. Montazer
Figure 2 Learners grouping sub-system architecture (see online version for colours)
Modules of this subsystem are as follows:
− Learner model database: All Information regarding characteristics describing every
learner are stored in this database, so that it can be used when needed to identify
group structures or classify learners into identified classes.
− Learner class database: The classes identified by group structure identification
process and learners in each group are stored in this database. Contents of this
database are used when needed to train the classifiers for classification new learners.
− Learner model agent: The task of this agent is to model each learner based on
his/her behaviour in the network. The learner model signifies the characteristics
affecting the determination of similarities between learning procedures of different
learners. The presented system uses Felder-Silverman learning style to describe the
learner model. This learner model agent is a collaboration agent tasked with
interacting with learning management system and monitoring learners’ behaviour in
− Group structure identification subsystem: The task of this subsystem is to identify
the e-learners’ group structures based on the learning style characteristics. This
subsystem has a continuous function carried out dynamically throughout the learning
process; therefore, another task of this subsystem is to reidentify the classes after a
change in structure. Initially, this subsystem does not have access to any data about
group structures, so it uses a clustering approach to initiate the process. This paper
uses a novel clustering method called modified fuzzy grafting clustering to detect
different group structures. While being comprehensive, this method provides
memory and time complexities as low as basic clustering techniques. Figure 3 shows
the algorithm of this clustering method.
An e-learning environment based on fuzzy grafting 175
Figure 3 Fuzzy grafting clustering algorithm (see online version for colours)
This method first cluster the dataset by a number of simple clustering methods, each
capable of identifying different types of data. These methods are called basic grafting
methods and should be able to determine the clusters’ centres. To achieve an appropriate
link between clusters, structure of each cluster should be assessed in conjunction with
that of other clusters. Therefore, once clusters of each basic grafting clustering method
are obtained, they undergo a fuzzy modelling process, which yields a series of clusters
corresponding to the results of basic clustering methods. Then a fuzzy comparison
operator determines the optimal clusters in each correspondence.
After selecting the optimal clusters, duplicate and missing elements are allocated to
their appropriate clusters, so ultimately a fuzzy combination of several clustering
methods detects the group structures with greater precision. The greater diversity of basic
grafting clustering methods, i.e., their ability to detect different types of structures, will
enhance the accuracy of final clustering; therefore, this study uses K-means technique
and modified fuzzy snap-drift neural network as basic grafting methods. The modified
176 M.S. Rezaei and G.A. Montazer
fuzzy snap-drift neural network is itself a combination of other clustering method, which
are described below.
3.2.1 Modified fuzzy snap-drift neural network
The proposed neural network is a combination of fuzzy ART network architecture and
snap-drift learning process. The proposed network consists of three layers; the number of
neurons in first and second layers equals the number of inputs (the number of attributes
of input pattern) and the number of neurons in third layer is initially 1 but eventually
increases to the number of identified clusters. Neurons of first and second layers are
linked one-to-one, but the links between neurons of third and second layers are bilateral.
We use a sigmoid threshold function for neurons of second and third layers and a linear
one for neurons of the first layer (Figure 4).
Figure 4 Modified fuzzy snap-drift neural network architecture
The top-down weights are initialised with half of the range of learners’ attributes and the
bottom-up weights are determined using Eq. (1) based on the corresponding top-down
weights (Brown, Draganova and Lee, 2009).
wis initial value of top-down weights and Nis the number of input layers. The
input vector I enters to the network and the output is calculated in the second layer using
Eq.(2) (Brown, Draganova and Lee, 2009):
The neuron that has max value of
is selected as winner cluster. The output of the
winner neuron is initialised with 1 and the other neuron of F2 layer is initialised with 0
An e-learning environment based on fuzzy grafting 177
temporarily. To approve assignment of pattern to the cluster corresponding to winner
neuron, in the next step, relative similarity input pattern with selective cluster is
calculated using Eq. (3):
W−is the Euclidean norm of distance between input pattern with centre of winner
cluster (up-down weights) and
is the Euclidean norm of input pattern. The equation
measures relative similarity of input pattern and corresponding pattern of winner neuron.
If the measured value is less than that of accept threshold, the pattern is assigned to the
cluster. After assigning the input to the cluster, the top-down weights of corresponding
neuron should be updated. It is shown in Eq. (4) (Brown, Draganova and Lee, 2009;
Lee, Palmer-Brown and Roadknight, 2004).
new old old old
Ji Ji Ji Ji
=− ∩ + + − (4)
wis the current up-down weight between ith and jth neurons, p is the network
performance feedback index, I is the fuzzy input vector andβ is the learning rate constant.
So ∩operator is fuzzy intersection which is considered as product operator here. In low
performance, the p variable is assigned from value 0 and the learning function turns to
Eq.(5). The performance evaluation criterion is Davies-Bouldin index. This index is
calculated for each epoch. Therefore, if it is greater than 2.5, then the performance is
considered bad; otherwise, the performance is considered good,
In high performance, by setting p = 1 in Eq. (4), Eq. (6) will be included (Brown,
Draganova and Lee, 2009)
( ) () ()
new old old
Ji Ji Ji
After updating up-down weights, the bottom-up weights of network are updated by
w is the Euclidean norm of up-down weight of ith neuron
(Lee, Palmer-Brown and Roadknight, 2004).
3.3 Subsystem of learners’ assignment to identified groups
This subsystem is responsible for assigning learners to the identified groups in which a
process is considered in order to automatically detect changes in group structure. This
process, based on the expansion of non-convergence in categorising learners, determines
178 M.S. Rezaei and G.A. Montazer
the need to reidentify group structures. In this subsystem, quantum neural network
(QNN) is used for categorisation; since these types of networks are able to categorise
learners with the lowest computational complexity and most accuracy. The high accuracy
of these networks in pattern detection is caused by the expansion of parallel power of
processing in the networks (Purushothaman and Karayiannis, 1997). The utilised
quantum neural network is a three-layer network. This network includes the input ni, a
hidden layer involving nh multilevel neurons and no output neurons. The weight
connecting of the ith output neuron to the jth hidden layer neuron with wij and connecting
weight of the jth hidden layer neuron from the kth feature vector is defined in the
jk jk h jk j
For every k
, x0,k = 1 and βh are the slope factors.
n is the number of energy levels and
indicates the jump positions in transfer function. The input of the ith output neuron
from the kth input vector is described in the equations below:
ik ij jk
ˆsgm( ( )
ik o ik
For every k, 0, 1
and βo are the sloping factors of output transfer function. The output
of each neuron is expressed by the following equation:
In which, v is the network weight connecting vector,
is the input vector and sgm
represents the sigmoid function applied on them. Quantum intervals, which refer to the
weight level of the multilevel transfer functions, are determined with the parameter of
jump positions. They are representation of discrete localised cells in feature space,
which involve feature vectors whose ambiguity level’s approximate number are intended
as their membership degree in data set classes (Purushothaman and Karayiannis, 1997).
In order to learn the network, first, the weight connecting should be adjusted by the
standard backpropagation algorithm and quantum intervals of hidden layer neurons. The
output changes of the pth hidden layer neuron for the mth category of m
C is as follows:
pm pC pk
An e-learning environment based on fuzzy grafting 179
C is the Euclidean measure of .
C The adjustment of
performed according to minimisation of the objective function defined by the 2
all categories and all hidden layer neurons in the following form:
pm pC pk
== == ∈
∑∑ ∑∑ ∑
Consequently, the updated
equation will become:
Δ= − × −
qk qk qk
vh h=− (19)
αθ refers to the learning tone. Therefore, this network can be regarded as, i.e., first-layer
neuron and first- and second-layer weights are classic and only the second-layer neuron
follows the quantum computations.
3.4 Subsystem of content assignment to learning groups
This subsystem is responsible for assigning learning content according to rules extracted
from expert and instructor opinion. Such subsystem consists of two units of ‘creating
content assignment rules’ and ‘content presentation’ and database of ‘rules of content
assignment to learning groups’.
A - Unit of Creating Content Assignment Rules: This unit is responsible for
creating rules of content assignment based on the information entered by the instructor.
These rules express the appropriateness of each learning object with the learning style of
learning groups with qualitative terms (very low, low, medium, high and very high). For
example, if the learning object LO1 is suitable from expert perspective for learners with
learning style of very high intuitive in the perception dimension and high visual in the
input dimension, this unit states the rule associated with it as follows:
“If the learner’s learning style is very high intuitive in the perception dimension and
high visual in the input dimension, then the learning object LO1 is suitable to present to
B - Content Presentation Unit: This unit is responsible for determining appropriate
content stored in content management unit for each of the learner groups. In this article,
decision tree method is used to assign learning content for each group of learners. Each
group, at any level of decision tree, evaluates the condition related to one of Felder-
Silverman learning style dimensions to reach as suitable content for each group of
learners. In Figure 5, the term linguistic variable of the learning style is demonstrated in
the learner grouping results as numerical intervals.
180 M.S. Rezaei and G.A. Montazer
Figure 5 Learners group’s value in linguistic variable of learning style
According to Figure 5, the normalised amount of the cluster centre Euclidean measure for
each group of learners determines the counterpart of that group with each of the amounts
of learning style linguistic variables. Based on these counterparts, suitable learning
objects are selected to present to each group of learners in accordance with rules defined
C - Database of Content Assignment Rules to Learner Groups: The content
assignment rules to learner groups are stored in this database so as to be used during
decision-making to choose suitable content for learner groups.
4 Evaluation of proposed adaptive e-learning system
The efficiency of the proposed adaptive e-learning system, with the smart capacity for
grouping learners, was evaluated in order to assess its effectiveness in the improvement
of learning environment. For this, a web-based IT enterprise architecture course was held
in the form of the proposed learning adaptation system. The characteristics of the learners
and the course are presented in Table 1.
Table 1 Specifications of learners and learning course
Learners Number 40 learners
Average age 23.2 years
Standard deviation of age 1.5 years
Gender Male and female
Learning course Number of learning session 12
Number of learning concept 12
Number of learning object 56
Duration of course 6 weeks
An e-learning environment based on fuzzy grafting 181
In order to evaluate the effect of learner grouping and the proposed method in the
improvement of education process, the participants in this course were randomly divided
into four groups:
• The first group (Gw) comprised 10 learners who passed the course without the
presentation of learning adaptive services. In this group, the learning objects were
offered to learners merely by the pre-knowledge order and randomly. In fact,
education was not customised for them.
• The second group (GK-means) comprised 10 learners who passed the course as
customised and in the form of learning adaptation groups. The learners in this group
were categorised using the common K-means method and, subsequently, were
presented with suitable learning content based on their group.
• The third group (GFGCM_w) comprised 10 learners who passed the course as
customised and in the form of learning adaptation groups. In the course presented to
this group, the learners were categorised using the proposed method but without
activating the possibility of group modification (reidentification of group structures)
and, afterwards, were presented with suitable learning content based on their group.
• The fourth group (GFGCM) comprised 10 learners who passed the course as
customised and in the form of learning adaptation groups. In the course presented to
this group, the learners were categorised using grouping method and with the
possibility of group modification and, then, were presented with suitable learning
content based on their group.
To evaluate the impact of using the proposed system in the conducted course, the
following criteria are used: ‘academic success’, ‘academic satisfaction’ and ‘time of
presence in the system’ (which is complementary of academic satisfaction in system
assessment and learners’ satisfaction of it). Academic success observes the learning rate
of the learner from the presented concepts during the course and academic satisfaction
observes the satisfaction rate of the learner from being present in the e-learning
environment and the ease in learning concepts of the course.
4.1 Academic success
In Table 2, the mean and standard deviation score of various groups are depicted in a test
involving seven theoretical essay questions, three theoretical multiple choice questions
and four qualification essay questions.
Table 2 Average and deviation of different groups
Number of learners 10 10 10 10
Mean scores 8.90 13.22 15.63 16.25
Standard deviation of scores 3.26 1.32 1.02 0.98
In order to accurately compare groups with each other, unilateral analysis of variance
(St and Wold, 1989) is used and the differences between the results concerning academic
success criteria of the learners present in the groups are evaluated.
182 M.S. Rezaei and G.A. Montazer
Table 3 demonstrates the results of final test analysis of variance. According to these
results, it is observed that the score of group Gw learners (who are denied customisation)
is significantly different from the other groups. Furthermore, the score of group GK-means
learners indicates a significant difference from the third and fourth groups. The difference
between learning customisation for GK-means group and learning customisation for
GFGCM_w and GFGCM groups is merely limited to the grouping method used to determine
the comparative group of learners; consequently, it can be assumed that using the
proposed grouping method in adaptive learning leads to the improvement of learners’
academic success. In addition, it is viewed that there is not a significant difference
between the learners’ scores in GFGCM_w group and GFGCM group. As a result, although the
learning style of some learners changes during learning in GFGCM group, this feature does
not lead to improving academic development of learners.
Table 3 Final test analysis of variance result
F statistic p-Values (p < 0.05) Significant difference
Between GK-means and Gw groups 7.6 3.18 Yes
Between GFGCM_w and Gw groups 8.8 3.18 Yes
Between GFGCM and Gw groups 8.9 3.18 Yes
Between GK-means and GFGCM_w
5.6 3.18 Yes
Between GK-means and GFGCM
4.4 3.18 Yes
Between GFGCM and GFGCM_w
3.12 3.18 No
4.2 Academic satisfaction
In order to assess learners’ satisfaction from various forms of learning personalisation, at
the end of each session, a questionnaire involving four questions were presented to the
learners. These questions contain five choices. Its question and objective are shown in
Table 4 Academic satisfaction questions
Question number Text of question Goal of question
1 How much did you learn in this
Learners grouping, determination
of learning style change
2 Was format of lesson’s content
Determination group of learners
in input dimension,
determination of learning style
3 How much do you satisfy about
Determination group of learners
in perception, understand and
determination of learning style
4 Would you like to use the
An e-learning environment based on fuzzy grafting 183
The mean scores of each learner for each question are provided in Table 5. These
amounts for the groups have been evaluated, using one-way variance of analysis
statistical method, the results are demonstrated in Table 6. The consequent results
indicate that, all in all, grouping of learners and presenting learning content suitable to
their adaptive group has effectively enhanced their satisfaction. On the other hand, in an
environment where the proposed method is used to group the learners, the learners’
satisfaction has had a significant difference compared with the other groups. The effect of
detecting change in learning style during learning and regrouping of learners has been
positive in the expansion of their academic satisfaction. These results are seen in three
out of four questions on academic satisfaction measurement.
Table 5 Result of academic satisfaction questionnaire
number Index Gw G
1 Mean 1.7 3.0 3.8 3.9
Standard deviation 0.95 0.77 0.77 0.87
2 Mean 2.5 2.9 3.0 3.0
Standard deviation 1.18 0.70 0.70 0.68
3 Mean 2.5 3.4 4.0 4.0
Standard deviation 1.65 0.82 0.65 0.68
4 Mean 2.2 3.54 3.7 3.7
Standard deviation 1.40 0.82 0.88 0.96
Table 6 Analysis of variance result of academic satisfaction
number Examined goal
(p < 0.05)
1 Learners grouping GK-means, Gw 5.3 3.18 Yes
Determination group of
GK-means , GFGCM_w 12.5 3.18 Yes
learning style change
GFGCM, GFGCM_w 7.2 3.18 Yes
2 Learners grouping GK-means, Gw 3.9 3.18 Yes
Determination group of
GK-means , GFGCM_w 3.2 3.18 No
learning style change
GFGCM, GFGCM_w 2.6 3.18 No
3 Learners grouping GK-means, Gw 15.6 3.18 Yes
Determination group of
GK-means, GFGCM_w 6.6 3.18 Yes
learning style change
GFGCM, GFGCM_w 5.6 3.18 Yes
4 Overall satisfaction GK-means, Gw 9.25 3.18 Yes
Overall satisfaction GK-means , GFGCM_w 5.6 3.18 Yes
Overall satisfaction GFGCM, GFGCM_w 4.2 3.18 Yes
184 M.S. Rezaei and G.A. Montazer
4.3 Presence in the system
The time of learners’ presence in the system is one of the indicators which can signify its
effectiveness in case of learners’ satisfaction with an e-learning environment.
Consequently, the mean useful time when the learner interacts with the system is
calculated and demonstrated in Table 7. As is clear from the results, on average, learners
of the groups GK-means , GFGCM_w and GFGCM have the most useful time of presence in the
system, which, based on the affirmation of learners’ academic satisfaction in these
systems, indicate that the grouping of learners and the presentation of suitable learning
for each group lead to adapting the provided services with the learners’ needs and
encouraging learners to be present in the system. Among these three groups as well, the
participants of groups GFGCM_w and GFGCM, for the grouping of which the proposed
method has been used, have spent more time than the participants of GK-means group (for
which the grouping method of K-means has been used). This means that the satisfaction
level in the learners of these groups is higher than GK-means group and the proposed
method has been able to provide a more absorbing environment for learners.
Table 7 The time of learners’ presence in the system
Total learners Learners finishing the course
Number of learners Average (hours) Number of learners Average (hours)
Gw 10 6.1 8 9.2
GK-means 10 22.3 10 22.3
GFGCM_w 10 30.6 10 30.6
GFGCM 10 27.2 10 27.2
In this paper, a new adaptive e-learning system is introduced for content adaptation to
learners. This system has the capacity to automatically group learners based on their
learning style, and a new method is used to enhance the accuracy of grouping learners.
The capacities of the proposed adaptive e-learning system include automatic grouping of
learners based on the identification of their group structures, and modification of the
generated groups through automatic categorisation possibility of new learners and those
learners whose traits have altered in the identified groups lead to differentiating this
system compared with other learning adaptation systems. The effect of these capacities in
improving the e-learning environment is examined by applying the proposed adaptive
e-learning system in an e-learning course. The results of the evaluation indicate that this
system has a positive effect on improving the learning environment. The proper accuracy
and velocity of this system and all of its automation in grouping has made the learning
process easy and absorbing; as a result, it has led to academic development and learners’
It seems that focus on the automation of personalisation instruments and the
customisation of an e-learning environment, by considering optimal accuracy and
velocity, can lead to performance improvement and learner motivation in an e-learning
environment. This capacity can improve the main disadvantage of these environments
An e-learning environment based on fuzzy grafting 185
compared with traditional learning environment in the absence of automatic adaptation of
the teaching process with the feedbacks from the learner’s learning style alterations.
Despite the efforts made in this article to apply learners’ learning style changes in their
grouping modification, the evaluation results do not reveal a significant improvement of
the learners’ academic satisfaction and development compared with a system without this
capacity. This can be caused by factors such as weakness in accurately identifying
changes and/or the inadequacy of content diversity for various styles in presenting to
them and it is recommended to be investigated in future works.
Akbulut, Y. and Cardak, C.S. (2012)‘Adaptive educational hypermedia accommodating learning
styles: a content analysis of publications from 2000 to 2011’, Computers & Education,
Vol. 58, No. 2, pp.835–842.
Bachari, E.E, Abelwahed, E.H. and Adnani, E.M. (2011) ‘E-learning personalization based on
dynamic learners’ preference’, International Journal of Computer Science and Information
Technology, Vol. 3, pp.200–216.
Botsios, S., Georgiou, D. and Safouris, N. (2008) ‘Contributions to adaptive educational
hypermedia systems via on-line learning style estimation’, Educational Technology & Society,
Vol. 11, No. 2, pp.322–339.
Brown, D.P.,Draganova, C. andLee, S.W. (2009) ‘Snap-drift neural network for selecting student
feedback’, International Joint Conference on Neural Networks, Atlanta, GA, pp.391–398.
Carver Jr., C.A., Howard, R.A. and Lane, W.D. (1999) ‘Enhancing student learning through
hypermedia courseware and incorporation of student learning styles’, IEEE Transactions on
Education, Vol. 42, No. 1, pp.33–38.
Chen, C.C., Chiu, P.S. and Huang, Y.M. (2015) ‘The learning style-based adaptive learning system
architecture’, International Journal of Online Pedagogy and Course Design (IJOPCD),
Vol. 5, No. 2, pp.1–10.
Dominic, M., Xavier, B.A. and Francis, S. (2015) ‘A framework to formulate adaptivity for
adaptive e-learning system using user response theory’, International Journal of Modern
Education and Computer Science (IJMECS), Vol. 7, pp.23–30.
Essalmi, F., Ben Ayed, L.J. Jemni, M. and Graf, S. (2010) ‘A fully personalization strategy of
e-learning scenarios’, Computers in Human Behavior, Vol. 26, No. 4, pp.581–591.
Gilbert, J.E. and Han, C.Y.(2002) ‘Arthur: a personalized instructional system’, Journal of
Computing in Higher Education, Vol. 14, No. 1, pp.113–129.
Jin, D.,Qinghua, Z. Jiao, D. and Zhiyong, G. (2006) ‘A method for learner grouping based on
personality clustering’, 10th International Conference on Computer Supported Cooperative
Work in Design (CSCWD’06, 2006), IEEE, pp.1–6.
Lee, S.W., Palmer-Brown, D. and Roadknight, C.M.(2004) ‘Performance-guided neural network
for rapidly self-organising active network management’, Neurocomputing, Vol. 61, pp.5–20.
Minetou, C.G. and SYC, X.L.(2005) ‘Grouping users’ communities in an interactive web-based
learning system: a data mining approach’, Fifth IEEE International Conference on Advanced
Learning Technologies (ICALT’05), Santo, Dominican Republic.
O’Donnell, E., Lawless, S., Sharp, M. and Wade, V.P. (2015) ‘A review of personalized e-learning:
towards supporting learner diversity’, International Journal of Distance Education
Technologies, Vol. 13, No. 1, pp.22–47.
Papanikolaou, K.A., Grigoriadou, M., Kornilakis, H. and Magoulas, G.D. (2003) ‘Personalizing the
interaction in a web-based educational hypermedia system: the case of inspire’, User
Modeling and User-Adapted Interaction, Vol. 13, No. 3, pp.213–67.
186 M.S. Rezaei and G.A. Montazer
Peter, S.E.,Bacon, E. and Dastbaz, M. (2010) ‘Adaptable, personalized e-learning incorporating
learning styles’, Campus-Wide Information Systems, Vol. 27, No. 2, pp.91–100.
Purushothaman, G. and Karayiannis, N.B. (1997) ‘Quantum neural networks (QNNs): inherently
fuzzy feedforward neural networks’, Neural Networks, IEEE Transactions, Vol. 8, No. 3,
Sonwalkar, N. (2015) Cloud Based Adaptive Learning Systems, White Paper http:// static1.
Specht, M. and Oppermann, R. (1998) ‘ACE-adaptive courseware environment’, New Review of
Hypermedia and Multimedia, Vol. 4, No. 1, pp.141–161.
St, L. and Wold, S. (1989) ‘Analysis of variance (ANOVA)’, Chemometrics and Intelligent
Laboratory Systems, Vol. 6, No. 4, pp.259–272.
Triantafillou, E., Pomportsis, A. and Demetriadis, S. (2003) ‘The design and the formative
evaluation of an adaptive educational system based on cognitive styles’, Computers &
Education, Vol. 41, No. 1, pp.87–103.
Zakrzewska, D. (2009) ‘Cluster analysis in personalized e-learning systems’, Intelligent Systems for
Knowledge Management, Vol. 252, pp.229–250.
Zhang, K., Cui, L., Wang, H. and Sui, Q. (2007) ‘An improvement of matrix-based clustering
method for grouping learners in e-learning’, 11th International Conference on Computer
Supported Cooperative Work in Design, 2007 (CSCWD, 2007), IEEE, pp.1010–1015.