Int. J. Computer Aided Engineering and Technology, Vol. 11, Nos.

4/5, 2019 495

Advanced prediction of learner’s profile based on

Felder-Silverman learning styles using web usage
mining approach and fuzzy c-means algorithm

Youssouf El Allioui
Laboratoire des Sciences des Matériaux,
des Milieux et de la Modélisation (LS3M),
Hassan First University,
25000, Khouribga, Morocco
Email: youssouf.elallioui@uhp.ac.ma

Abstract: Automatic prediction of learner’s profile is an important requirement

for personalised e-learning. This can be provided based on the learning
behaviours of the learners. In this work, the learning behaviour is captured
using the web usage mining technique, preprocessed and converted into the
XML format based on sequences of accessing contents. These sequences are
mapped to the eight categories of Felder-Silverman learning style model
(FSLSM) using fuzzy c-means (FCM) algorithm. A gravitational search based
back propagation neural network (GSBPNN) algorithm is used for the
prediction of learning styles of a new learner. In this algorithm, the neural
network approach is modified by calculating the weights using gravitational
search algorithm. The accuracy of the prediction model is compared with the
basic back propagation neural network (BPNN) algorithm. The result shows
that the captured data is labelled as per FSLSM and the accuracy is more in
GSBPNN as compare to BPNN.

Keywords: learner’s profile; Felder-Silverman learning styles; web usage

mining; WUM; fuzzy c-means algorithm.

Reference to this paper should be made as follows: El Allioui, Y. (2019)

‘Advanced prediction of learner’s profile based on Felder-Silverman learning
styles using web usage mining approach and fuzzy c-means algorithm’, Int. J.
Computer Aided Engineering and Technology, Vol. 11, Nos. 4/5, pp.495–512.

Biographical notes: Youssouf El Allioui received his PhD in Computer

Science from the Sidi Mohamed Ben Abdellah University of Fez, Morocco. He
is an Associate Professor at the Department of Computer Science, Hassan 1st
University, Morocco. His research interests include issues related to semantic
web, knowledge management, web personalisation and recommender systems,
e-learning and semantic e-learning. He has published several research papers at
international journals and conferences. He has served as a member of the
technical committees of several international conferences and workshops.

Copyright © 2019 Inderscience Enterprises Ltd.

496 Y. El Allioui

1 Introduction

Adaptive e-learning system enhances the efficiency of online education by providing

personalised contents and user interfaces. This can be changed with respect to learner’s
requirements that are captured through the usage patterns. Personalised environment
should be adaptive and should change according to the learner’s requirements. This
usually differs among learners who attend the same course. In order to understand the
learner’s requirements, learners with the similar interests can be captured based on the
usage data of each learner. The similar sequence of learners is grouped together based on
the frequency of access and time spent on the learning materials. The similar interests
(usage data) and similar sequences are available in the form of learning objects on the
e-learning portal. The similar sequences of learners are considered for grouping the
learners based on Felder-Silverman learning style model (FSLSM) categories.

Figure 1 Dimensions and categories of FSLSM (see online version for colours)

The learning behaviour can be mapped to learning styles by using suitable learning style
model. The FSLSM is the most popular learning style model available which emphasises
on various categories of learners. The FSLSM defines four dimensions (pre-processing,
perception, input and understanding) and eight categories of learners (active, reflective,
sensing, intuitive, visual, verbal, sequential and global), which are shown as in Figure 1.
To make traditional e-learning system adaptive, there is a prime necessity to integrate
the learning styles in to the portal and provide contents to the learners as per learning
styles. Two important steps to integrate the learning styles into adaptive e-learning
system are mentioned as follows:
1 Identify the learning style using captured data (sequences).
2 Modify the e-learning portal as per identified learning styles.
 Advanced prediction of learner’s profile based on Felder-Silverman 497

The learning styles can be identified based on standard learning style model such as
FSLSM. The captured data can be mapped to learning style categories of FSLSM, which
is possible through clustering and classification techniques (Abdullah et al., 2015a). To
perform the clustering and classification of learner’s sequences following research
questions need be accentuated.
1 What are the various parameters/feature values that need to be considered for each
type of learner as per FSLSM categories?
2 How to group the captured learner’s sequences into FSLSM categories in order to
assign labels to each type of learner?
3 How to classify the new learner’s sequence into defined FSLSM categories of the
learners?
The above mentioned research questions can be answered as follows:
1 The learning styles of learners are identified based on defined parameters/feature
values of FSLSM.
2 The available sequences of learners need to be labelled as per eight categories of
FSLSM learners.
The clustering and classification algorithms can be used to identify the learning styles.
This can be achieved by grouping the available sequences and classifying web usage data
of a specific learner.
Rest of the paper is organised as follows. Section 2 explains the web usage mining
(WUM) technique. Section 3 discusses related research work in the area of clustering and
classification techniques. Section 4 discusses about the learning components of e-learning
portal. Section 5 describes the methodology used for clustering and classification of
learning sequences. Section 6 presents the results and related discussions and Section 7
mentions the concluding remarks.

2 Literature review

Abdullah et al. (2015b) have proposed a new approach to classify students dynamically
depending on their learning style. This approach is experimented on 35 students for data
structure online course created using Moodle. The learning style of each student is
identified according to FSLSM by extracting student’s behaviour and data from Moodle
log. Also, learning styles based on the behaviour have been compared with quiz results
conducted at the end of the course. The e-learning portal considered for study is Moodle
and it is static in nature. The authors have not considered the learning components of
portal such as pages and files, which are useful to define the learning styles based on
FSLSM.
Feldman et al. (2015) have reviewed the current trends in the field of automatic
detection of learning styles. They have presented the analysis result of different
techniques and discussed some limitations, implications and research gaps. Authors have
described various detection techniques, which are mainly categorised under classification
mechanism of data mining. Authors have also proposed the user model for the education
system that is used to detect the learning styles.
498 Y. El Allioui

Chang et al. (2009) have proposed a learning style classification mechanism to

classify and identify student’s learning styles. The mechanism modifies k-nearest
neighbour (k-NN) classification algorithm and combines it with genetic algorithms
(GAs). The learning behavioural features of 117 elementary school students are collected
and then classified. The experimental results indicate that the classification mechanism
can effectively classify and identify student’s learning styles. The mechanism has not
considered the specific features of learning styles based on a particular learning style
model. The selection of learning style model influences the characteristics of mechanism.
Authors have not considered the use of learning style model for the improvement of
e-learning system with the recommendation mechanism based on learning styles of
learners.
Abdullah et al. (2015b) have discussed about the impact of learning styles in the
e-learning environment. The naïve-Bayes tree (NBTree) classification technique is used
to classify the learning style of the students as per FSLSM. In this approach the learning
styles are captured using questionnaire index of FSLSM, which has its limitations to
consider the learner’s online usage behaviour.
Latham et al. (2013) have combined the multi-layer perceptron and artificial neural
network (ANN) to classify the learner’s data. The learner’s data is captured based on
FSLSM questionnaire index and considered four categories of learners (active/reflective
and sequential/global) for classification.
Al-Azawei and Badii (2014) have discussed the importance of integrating the
learning styles into educational hypermedia systems (EHS). Authors have also discussed
various approaches such as Bayesian network, decision tree and ANN, etc. to classify log
data accurately.
Bernard et al. (2015) have proposed an approach that uses ANN to identify student’s
learning styles. Authors have considered the FSLSM questionnaire approach for learning
styles of students. Students who had spent more than 5 minutes to fill the questionnaire
have been stored in the database. This approach can be applied dynamically to detect the
learning styles.
Takács et al. (2011) have proposed a method to make learning more efficient.
Authors have tried to adapt the process of learning to learner’s individual needs.
Learner’s characteristics such as learning preferences, sensual types, learning styles and
motivation are identified as part of the proposed work. Authors have used questionnaire
approach to collect the data. The results of this methodology are analysed with clustering
algorithms, decision trees and principal component analysis (PCA). Authors have also
analysed that learners cannot be divided into various groups that possess similar
properties and it is not possible to deal with every learner individually.
Truong (2016) has reviewed various works that has been tried to integrate the
learning styles and adaptive e-learning system. He has also discussed about various
aspects of learning style theories, real-time learning style predictors, learning style
classifications algorithms and adaptive learning system using learning style applications.
In learning style theories author has listed Felder and Silverman model, kinaesthetic and
Kolb’s learning style model, Honey and Murnford’s learning style model, etc. In
real-time learning style predictors, the main sources of data and the corresponding
attributes such as log files, user history, background knowledge, intelligent capability,
cognitive traits, motivation level, etc. have been discussed. In learning style classification
the author has explained about rule-based, Bayesian network and hidden Markov chain.
 Advanced prediction of learner’s profile based on Felder-Silverman 499

Hogo (2010) has presented the use of different fuzzy clustering techniques such as
fuzzy c-means (FCM) and kernelised FCM to find the learner’s categories and predict
their profiles. Fuzzy clustering reflects the learner’s behaviour more than crisp clustering.
The author have mentioned three different types of learners such as bad, worker and
regular students and not considered any standard learning style model for learning styles
of the learners. The usage data of students is captured based on questionnaire approach of
FSLSM.
Do (2015) has proposed the use of gravitational search algorithm (GSA) in neural
network (NN) and compared with back propagation neural network (BPNN) for
classification. Xu and Zhang (2014) combined the GSA with NN and tested on different
data sets for classification.
Thiyagarajan et al. (2011) have discussed about the data mining techniques which can
improve the learning content organisation and learning objects recommendations.
Authors have focused on clustering and classification methods to predict the learning
styles of the learners. The learning behaviours have captured on Moodle portal, which is
static in nature and does not support dynamic recommendations. Also, the approach is
implemented for Kolb learning style.
Deborah et al. (2015) have identified FSLSM as a suitable model for learning style
prediction, especially in web environments. The authors have proposed to use fuzzy rules
to handle the uncertainty in the learning style predictions. However, this work has not
considered all the learner categories of FSLSM. The fuzzy rules are defined only on three
parameters such as mouse movement, document length and image area.
Markowska-Kaczmar et al. (2010) have discussed about providing individual learning
path through recommendation system in e-learning. Authors have used clustering and
classification for defining the learning profiles but not considered any specific learning
style model for learning behaviours.

3 Web usage mining

WUM is the third category in web mining. Indeed, as shown on Figure 2, the web mining
is of three types:
• Web content mining: which is the process of extracting useful information from the
contents of web documents
• Web structure mining: is the process of discovering structure information from the
web
• WUM: is the process of discovering meaningful patterns from data generated by
client-server transactions stored in web logs.
WUM refers to the automatic discovery and analysis of patterns in click streams, user
transactions and other associated data collected or generated as a result of user
interactions with web resources on one or more websites (Cooley et al., 1997; Mobasher,
2006; Srivastava et al., 2000). The goal is to capture, model and analyse the behavioural
patterns and profiles of users interacting with a website. The discovered patterns are
usually represented as collections of pages, objects, or resources. They are frequently
accessed or used by groups of users with common needs or interests.
500 Y. El Allioui

Figure 2 The general relationship between the categories of web mining (see online version
for colours)

WUM contribute to finding significant educational knowledge. It plays a vital role in the
personalisation aspects of any domain. In the advance of e-learning, the importance of
WUM grows larger than before. In this work, I use the WUM technique to capture the
learning behaviours of learners and converted it into the XML format based on sequences
of accessing contents on the e-learning portal. Finally, I make use of FCM algorithm to
map the sequences to the eight categories of FSLSM.

4 Learning components of e-learning portal

Following are the learning components defined for the e-learning portal.
• Announcement: instructor can convey any important instructions or suggestions
through this component. Learner can view by clicking on announcement link.
• Email: learner can send mail to instructor or any other learner in order to clarify
some doubts or share some materials.
• Discussion forum: learner can start discussions on specific topics or can get the
doubts solved online with other learners.
• Assignments: learner can read the assignment question and upload the completed
assignment for assessment.
• References: learner can see other important links or material by clicking on reference
link of each topic.
• Exercise: learner can solve multiple-choice questions on specific topic after going
through the contents.
• Topic list: learner can view the list of topics, which has been made available in the
portal.
 Advanced prediction of learner’s profile based on Felder-Silverman 501

• My account: learner can modify his/her personal details in the portal.

• Pre/main/advanced: topics have been arranged in the portal in three different levels
so that learners can go through any of these levels of the topics based on his/her
understanding.

4.1 Learning objects labelling as per FSLSM

The e-learning portal which is a combination of learning components called ‘pages’ and
learning contents called ‘files’. Each learning component type of contents in the portal is
considered as learning objects. To map learning objects onto FSLSM categories, all
learning objects need to be labelled depending on the preferences of each learner.
mapping of learning objects on FSLSM’s categories are shown in Table 1. These mapped
learning objects are considered as feature values for the clustering algorithm to label the
sequences.
Table 1 Learning objects as per FSLSM

Active learners Videos, PPTs, demo, exercise, assignments.

Reflective learners PDFs, PPTs, videos, announcements, references.
Sensing learners Examples, PDFs, videos, practical material.
Intuitive learners PDFs, PPTs, videos, forum, topic list, references.
Visual learners Images, charts, videos, references.
Verbal learners PDFs, videos, email, announcements.
Sequential learners Exercise, references, assignments, sequential.
Global learners Topic lists, references, exercise, assignment.

4.2 Characteristics of web pages as per FSLSM

The learners access various web pages on the e-learning portal. The web pages contain
following different learning objects.
• Learning contents in the form of text, ppt, videos, etc.
• Different topic links, forum, email, index of the sequential contents, profile details,
etc.
• The text contents such as various images, charts, graphs, diagrams, flowcharts, video
links, etc.
• Assignment with upload and download option for the answers, multiple choice
questions and search option.
The characteristics of the web pages are defined based on the above points, which can
map to FSLSM categories of learners. The e-learning system captures the usage access of
all the learners along with the details of the pages and files accessed by each learner. The
captured details of the pages are used to identify the learning styles of learners.
502 Y. El Allioui

5 Methodology

The approach of labelling the sequences and classification of new sequence is shown in
Figure 3. The captured sequences are labelled according to eight categories of FSLSM.
The FSLSM mapped learning objects are considered as feature values for labelling. Also,
the characteristics of web page are identified based on feature values of FSLSM. These
characteristics are used to understand the learning styles of the learner who is accessing a
specific web page. Around 1,235 similar sequences are grouped into eight clusters of
FSLSM. After clustering the existing sequences, the classifier model is trained using
gravitational search-based NN algorithm for the existing sequences. This algorithm
classifies the new sequence into FSLSM’s eight categories based on the calculated score
value. To validate the performance of the classification algorithm, k-fold cross-validation
algorithm is used (Agbonifo, 2013).

Figure 3 Learning style identification approach (see online version for colours)

The identified sequences of learners include the sequence of pages and files accessed by
each learner along with the time spent on each page and file. These sequences are not
labelled and not mapped to learning objects of FSLSM’s categories. To classify the new
sequence of any learner into any of the eight defined categories of FSLSM, all the
existing sequences of sequence file should be labelled or grouped based on the feature
values. Hence, to identify the learning styles of the learner and to category of learner,
clustering and classification approaches are used. The proposed approach uses two
algorithms.
 Advanced prediction of learner’s profile based on Felder-Silverman 503

• FCM algorithm to label or map the sequences into FSLSM categories.

• Gravitational search based back propagation neural network (GSBPNN) algorithm to
classify the new sequence of existing or new learner as per FSLSM categories.

5.1 FCM algorithm

The FSLSM’s feature values and the sequences of learners are given as an input to the
FCM algorithm. Based on the feature values, the sequences are grouped together. The
objective function of the FCM is calculated as shown in equation (1) (Joshi et al., 2011;
Agbonifo, 2013).
n c
Jm =  Mi =1 j =1
m 2
ij d ij (1)

where distance between cluster and sequence value is calculated as shown in equation
(2).
dij = Fi − C j (2)

The FCM technique is entrusted with the task of assigning membership to each and every
sequence values, which are linked to each singular cluster centre. If the sequence values
are very near to the cluster centre, its membership moves further toward the precise
cluster centre. The sum of all the membership values of each sequence should be equal to
one. Further, membership and cluster centres are revised according to the equation after
each iteration. After finding the distance value, the value is compared with the threshold
value. If the distance value is less than the threshold value then learner’s sequence is
grouped in that cluster. Here more than one cluster can have the same sequence value.
The step-by-step procedure is shown in Algorithm 1.

Algorithm 1 FCM algorithm

Initialise m = real number greater then 1

Mij = degree of member ship function xi in cluster j
Fi = input sequence
Cj = centre of the cluster
∈ = threshold value
Step 1 Initialise the membership function
M = [Mij], M(0)
Step 2 Calculate the centre value
C = Cj, with M(K)
K


n
M ijm Fi
i =1
Cj =

n
M ijm
i =1

Step 3 Update the membership function MK and MK + 1

504 Y. El Allioui

1
M ij = 2
 Fi − C j  m −1

C

K =1 F − C

 i k 

Step 4 If ║MK+1 – MK║ < ε then stop, else go to Step 2.

In FCM algorithm, the learner’s learning style is grouped based on the centre value
selection. The centre values are calculated based on the feature values and assigned to
each cluster. Finally, all the sequences are grouped into eight clusters of FSLSM
categories and each sequence is assigned the label as per FSLSM. The flowchart of the
FCM algorithm is shown in Figure 4.

Figure 4 The flow-chart of the FCM algorithm

5.2 Goodness of clusters

5.2.1 Jaccard index
The clustering algorithm should form high quality clusters with low similarity between
the clusters. The Jaccard index is used to check the similarity between the clusters
 Advanced prediction of learner’s profile based on Felder-Silverman 505

(Zakrzewska, 2008). The steps of Jaccard index are shown in Algorithm 2. The FCM
algorithm results in the grouping of some sequences into more than one cluster. Hence, it
is necessary to have the distance between the clusters closer to one based on Jaccard
index algorithm.
Algorithm 2 Jaccard index algorithm

Initialise cluster of sequences

Step 1 Find unique sequences of each of the cluster.
Step 2 For each cluster calculate Jaccard index.
Ci ∩ C j
Jaccindex (Ci , C j ) =
Ci ∪ C j
Step 3 If Jaccindex(Ci, Cj) is equal to 1, the clusters are exactly similar otherwise different.
Step 4 Compute the average Jaccard index value to validate all the clusters.

5.2.2 Xie-Beni index

The Xie-Beni (XB) index defines the inter-cluster separation as the minimum square
distance between cluster centres and the intra-cluster compactness as the mean square
distance between each data object and its cluster centre. To identify the goodness of
intra-cluster similarity (Zakrzewska, 2008) XB indexing is applied and checked the
similarity of clustered points of FSLSM eight categories.
The XB index calculated using equation (3).

 
C N 2
μM Z k − Vi
i =1 k =1 ik
X (Z , U , V ) = (3)
(
C.min i ≠ j Vi − V j
2
)
where X = XB index, Z = centre value, U = member function, V = centre values,
μ = mean value.

5.3 GSBPNN algorithm

After labelling the sequences in the FCM algorithm, all the sequences are used to train
the classification algorithm. In order to provide the adaptive user interface, the new
sequence of learner’s session needs to be classified as per eight classes of FSLSM. To
classify the sequences, GSBPNN algorithm is used. The new sequence of each learner is
given as input to NN. The gravitational search is used to improve the accuracy of the
basic NN classification algorithm.

5.3.1 Gravitational search algorithm

GSA is a heuristic optimisation algorithm, which is based on Newton’s law of gravity
and interactions between the mass. The algorithm consists of number of agents, which are
selected randomly or based on specific condition. These agents are considered as
communicating objects through the force of gravity. The positions and masses of the
objects represent the solution to the specific problem. The iterations of GSA are based on
the change in positions. The change in the positions is defined using the fitness function,
506 Y. El Allioui

velocity and acceleration. These parameters change in every iteration of GSA to provide
better weight values (Xu and Zhang, 2014; Sabri et al., 2013). The steps are shown in
Algorithm 3.
Algorithm 3 Gravitational search algorithm

Initialise
Step 1 Initialisation of agents: initialise the positions for N number of agents.
X i = ( xi1 , …, xid , … xin ) for i = 1, 2, …, N

xid is the position of ith agent in the dth dimensions

Step 2 Compute the fitness function for each agent.
fitj (t) represents the fitness value of the jth agent in the iteration t.
best(t) and worst (t) are given below for maximisation or minimisation of fitness at iteration t.
best(t) = min(fitj(t)) AND worst(t) = max(fitj(t)) where j = 1 to N
OR
best(t) = max(fitj(t)) AND worst(t) = max(fitj(t)) where j = 1 to N
Step 3 Compute the gravitational cost.
 Itr 
G (t ) = G0 1 − where G0 is constant
 Total Itr 
Step 4 Compute the gravitational mass.
fit (t ) − worst (t )
mk =
best (t ) − worst (t )

mk
M i (t ) =

n
mk
k =1

Step 5 Calculate the acceleration.

F (t )
ai (t ) = where F = total force
M i (t )
Step 6 Update the agent velocity and positions.
vi(t + 1) = rand ∗ vi(t) + ai(t) and xi(t + 1) = xi(t) + vi(t + 1)

5.3.2 NN algorithm
NN is used for processing of information that is based on nervous system of biological
process. The component of this process is brain, neurons and connection between them
for information process. NN consists of large number of neurons, connected with each
other to solve the specific problem. The applications of NN are classification, pattern
recognition, image compression, prediction, etc. NN works on the learning process,
which involves the adjustments of weights, which are assigned to the connections
between the neurons.
NN is organised in layers, where layer consists of many interconnected nodes. These
interconnected nodes are mimicking the characteristics of biological neurons by
processing the input data. The output of each node is called as the activation or node
values. The connection between the nodes is represented by weights, which could be
 Advanced prediction of learner’s profile based on Felder-Silverman 507

positive or negative. These weight values decide how input data are related with the
output data. Weight values are determined at the time of training for each iteration
through the NN. Thus, the network learns to identify different classes based on the input
data characteristics. After the training procedure, the NN could be applied for the
classification of new data. The following steps are used to the trained NN.
• The network input nodes are activating by using data sources that match the data
sources used to train the network.
• The data is forwarded through the interconnected nodes in NN.
• The network output nodes are activated and it is decided the outcomes of the
algorithm.

5.3.3 Back propagation neural network

NN learns through the algorithm called back propagation. With back propagation, the
input data is repeatedly fed to the NN. Each time the output of the NN is compared to the
desired output and an error is computed. This error is then fed back (back propagated) to
the NN and used to adjust the weights. Hence, the error decreases with each iteration and
the neural model gets closer and closer to producing the desired output. This process is
known as training. After training the NN model, testing data checked for classification
(Özpolat and Akar, 2009). The steps of BPNN are described in Algorithm 4. The GSA
gives optimised weights, which can be used in the process of BPNN. The accuracy of
classification algorithm is based on the weights assigned the connections between the
neurons in the BPNN. To assign the optimised weights GSA is used. The BPNN consists
of two input nodes, one sequence node and one bias node, three nodes at one hidden layer
and eight output nodes for eight classes of FSLSM (Do, 2015). The BPNN algorithm
classifies the learner’s sequence into multiple classes, i.e., eight categories (classes) of
FSLSM.
Algorithm 4 BPNN algorithm

Initialise
Step 1 Initialise two nodes for the input layer, three nodes for the hidden layer and eight
nodes for the output layer.
Step 2 Activate uniti in the input layer Xi = ai(m) where i = 1, 2 input parameters and ai(m) is
the input vector.

Step 3 Activate unitj in the hidden layer X hj = W ( x ) + b

ij i where j = 1 to 3 (hidden layer)

1
Step 4 Output signal from unitj in hidden layer θ hj = h .
1 + e( − X j )

Step 5 Activate unitk in the output layer X 0k = W θ k

j h
j where k = 1 to 8 output classes.

 (d
Step 6 1
Mean square error for the kth output E = k − ok ) 2 where dk = actual output
n
AND ok = experimented output
n = number of patterns in the samples.
508 Y. El Allioui

5.4 Validation of algorithm

The classification algorithm is validated using k-cross fold validation technique. This
technique is commonly used for validating the performance of classifier by dividing the
data set into k cross folds. For given set of M sequences, one data set is used for training
and k-1 data sets are used testing the classifier. The steps are shown in Algorithm 5.
Algorithm 5 k-cross fold validation algorithm

Initialise
M = number of training sequences as a training set
k = number of folds
ni = number of sequences classified wrong in foldi
Step 1 Arrange the training sequence M as a data set.
Step 2 Divide the training sequences into k parts.
Step 3 For each i where i = 1, 2, …, k
Step 3.1 Train the classification algorithm using the all sequences except the foldi.
Step 3.2 Test the algorithm for all the sequences of foldi.
Step 3.3 Compute ni: wrongly classified sequences.


k
ni
Step 3.4 Compute the estimate E = i =1
.
m

5.5 Precision and recall

The accuracy of classification algorithms is measured based on the number of correct
predictions. To prove the validity of the predictions and robustness of the models,
accuracy alone is not sufficient. However, there is a requirement to calculate the
precision and recall for each class label and analyse the performance.
Precision (P): to find the number of input sequences which are correctly classified in
a given class. Precision is calculated as in equation (3).
TP
P= (3)
TP + FP
Recall (R): to find the number of sequences which are correctly identified in the given set
of sequences. Recall is calculated as in equation (4).
TP
R= (4)
TP + FN
Accuracy (A): to measure the performance by finding the ration of correctly predicted
sequences. Accuracy is calculated as in equation (5).
TP + TN
A= (5)
TP + TN + FP + FN
 Advanced prediction of learner’s profile based on Felder-Silverman 509

F-measure (F1 score): Since the class distribution is uneven, it is always better to
calculate the F1 score. The F1 score provides the weighted average of precision and
recall. Hence, it consider both false positive (FP) and false negative (FN).
2( R × P )
F1 = (6)
( R + P)

TP ≡ true positive number of sequences that are correctly classified to be positive.

TN ≡ true negative number of sequences that are correctly classified to be negative.
FP ≡ false positive number of sequences that are incorrectly classified as positive.
FN ≡ false negative number of sequences that are incorrectly classified as negative.

6 Results and discussions

6.1 Clustered sequences based in FCM algorithm

1,235 similar sequences are considered for clustering based on FCM algorithm. Centre
values for each cluster are computed based on the feature values, i.e., mapping of
learning objects to FSLSM categories. The result of clustering is shown in Table 2.
Table 2 Result of FCM algorithm

Clusters Active Reflective Sensing Intuitive Visual Verbal Sequential Global

No. of 169 144 138 162 256 133 144 152
sequences
Total 1,298

The sequences are labelled as per eight categories of FSLSM. The total numbers of
clustered sequences (1,298) are more than the input sequences (1,235) as some of the
sequences belong to more than one cluster based on the feature values.
The goodness of clusters is computed by applying Jaccard index algorithm. The
average Jaccard index computed for eight clusters is 0.72. The Jaccard index value is not
exactly 1, as some sequences are clustered into more than one cluster. The validity of
FCM also evaluated by calculating XB index value for all the clusters. The calculated
index value is 0.02. The minimum index value defines the best partitions of the clusters.
Based on the result of FCM, the sequences are labelled into FSLSM eight categories
for classification model.

6.2 Classification of sequence based on GSBPNN algorithm

The GSBPNN algorithm is used to classify the sequences as per the eight categories of
FSLSM. Once the model is trained based on sequences, which are labelled as per FCM,
the new sequence can be classified and learning styles of the learner identified. The
algorithm is executed for various numbers of iterations (I) and the accuracy obtained and
compared with BPNN algorithm, as shown in Table 3.
510 Y. El Allioui

Table 3 Result of classification algorithms

GSBPNN BPNN
I
A (%) P (%) R (%) F1 (%) A (%) P (%) R (%) F1 (%)
60 87.16 88.09 90.3 88.89 75.13 76.38 79.43 77.20
100 89.45 90.03 92.18 90.87 77.09 78.18 81.02 79.03
200 95.93 96.33 99.05 97.67 78.83 79.9 83.1 80.12
Notes: I ≡ Number of iterations
A ≡ Accuracy, is calculated as in equation (5)
P ≡ Precision, is calculated as in equation (3)
R ≡ Recall, is calculated as in equation (4)
F1 ≡ F-measure, is calculated as in equation (6).
According to values defined in Table 3, the accuracy of GSBPNN classification
algorithm is 95.93% for 200 iterations. If the number of iteration increases, the accuracy
may increase but algorithm will take more time for execution.

6.3 About results

Although this promising results, many directions can be opened up for future
improvement. First, I have shown that the proposed approach is useful for a particular
domain. However, it can be implemented in a multi-domain platform by adding a topic
extraction layer to automatically detect area of expertise of each expert. Second, I can add
a temporal factor for selecting useful experts. In fact, exploring time-dependent expertise
can leverage expert selection phase and recommendation accuracy.

7 Conclusions and future work

To enhance the efficiency of online course, adaptive e-learning systems (AES) is a

promising research area. A necessary requirement in this area is to identify the learning
styles of the learners based on standard learning style model, which is useful to provide
adaptive learning experiences. In this context, the work describes a methodology to
automatically detect and identify the learning styles of learners using web log analysis
approach.
The captured web usage data is pre-processed and converted into XML format to
identify the unique sequences of each learner as per their sessions. These unique
sequences are further identified as similar and non-similar sequences irrespective of
length and order based on time spent and frequency. The similar sequences are mapped to
eight categories of FSLSM based on the learning objects defined for each category. The
mapping has been achieved by applying FCM algorithm on the similar sequences and
clustered them into eight groups of FSLSM which are defined as active, reflective,
sensing, intuitive, visual, verbal, sequential and global. Some sequences are assigned
multiple labels according to the featured values of FSLSM learning objects. The output of
FCM algorithm is checked for goodness between the clusters. The sequences are further
used for classification by training the GSBPNN model. Applying k-cross fold mechanism
validates the classification algorithm. The model executed for different number of
 Advanced prediction of learner’s profile based on Felder-Silverman 511

iterations and compared accuracy with basic BPNN algorithm. Finally the sequences are
classified into multiple classes.
In future, I am working on to obtain complete learner’s information’s and provide
personalised environment. The model recommends those required activities and/or
resources that would favour and improve their learning process and integrate learning
styles into Adaptive e-learning system to assess the effect of adapting educational system.

References
Agbonifo, O.C. (2013) ‘Fuzzy c-means clustering model for identification of students learning
preferences in online environment’, International Journal of Computer Applications &
Information Technology, October–November, Vol. 4, No. 1, pp.15–21.
Al-Azawei, A. and Badii, A. (2014) ‘State of the art of learning styles-based adaptive educational
hypermedia systems (Ls-Baehss)’, International Journal of Computer Science & Information
Technology, June, Vol. 6, No. 3, pp.1–10.
Bernard, J., Chang, T-W., Popescu, E. and Graf, S. (2015) ‘Using artificial neural networks to
identify learning styles’, Artificial Intelligence in Education, June, Vol. 9112, pp.541–544.
Chang, Y-C., Kao, W-Y., Chu, C-P. and Chiu, C-H. (2009) ‘A learning style classification
mechanism for e-learning’, Computers & Education, September, Vol. 53, No. 2, pp.273–285.
Cooley, R., Mobasher, B. and Srivastava, J. (1997) ‘Web mining: information and pattern
discovery on the World Wide Web’, International Conference on Tools with Artificial
Intelligence (ICTAI-1997), IEEE Conference Publications, pp.558–567.
Deborah, L.J., Sathiyaseelan, R., Audithan, S. and Vijayakumar, P. (2015) ‘Fuzzy-logic based
learning style prediction in e-learning using web interface information’, Édité par Indian
Academy of Sciences, Sādhanā, April, Vol. 40, No. 2, pp.379–394.
Do, Q.H. (2015) ‘A hybrid gravitational search algorithm and back-propagation for training
feedforward neural networks’, Knowledge and Systems Engineering, Vol. 326, pp.381–392.
Feldman, J., Monteserin, A. and Amandi, A. (2015) ‘Automatic detection of learning styles: state
of the art’, Artificial Intelligence Review, August, Vol. 44, No. 2, pp.157–186.
Hogo, A.M. (2010) ‘Evaluation of e-learners behavior using different fuzzy clustering models: a
comparative study’, International Journal of Computer Science and Information Security,
March, Vol. 7, No. 2, pp.131–140.
Joshi, M., Vaidya, R. and Lingras, P. (2011) ‘Automatic determination of learning styles’, 2nd
International Conference on Education and Management Technology, IACSIT Press,
Singapore, pp.121–132.
Latham, A., Crockett, K. and Mclean, D. (2013) ‘Profiling student learning styles with multilayer
perceptron neural networks’, International Conference on Systems, Man, and Cybernetics,
IEEE, pp.2510–2515.
Abdullah, M.A., Alqahtani, A., Aljabri, J., Altowirgi, R. and Fallatah, R. (2015a) ‘Learning style
classification based on student’s behavior in Moodle learning management system’,
Transactions on Machine Learning and Artificial Intelligence, February, Vol. 3, No. 1,
pp.28-40.
Abdullah, M.A., Bashmail, R., Daffa, W. and Alzahrani, M. (2015b) ‘The impact of learning styles
on learner’s performance in e-learning environment’, International Journal of Advanced
Computer Science and Applications, Vol. 6, No. 9, pp.24–31.
Markowska-Kaczmar, U., Kwasnicka, H. and Paradowski, M. (2010) ‘Intelligent techniques in
personalization of learning in e-learning systems’, Computational Intelligence for Technology
Enhanced Learning, Studies in Computational Intelligence, Vol. 273, pp.1–23, Édité par
Springer, Heidelberg, Berlin.
512 Y. El Allioui

Mobasher, B. (2006) ‘Web usage mining’, Encyclopedia of Data, Idea Group, DePaul University,
USA.
Özpolat, E. and Akar, G.B. (2009) ‘Automatic detection of learning styles for an e-learning
system’, Computers & Education, September, Vol. 53, No. 2, pp.355–367.
Sabri, N.M., Puteh, M. and Mahmood, M.R. (2013) ‘A review of gravitational search algorithm’,
International Journal Advance Soft Computer Applications, November, Vol. 5, No. 3,
pp.1–39.
Srivastava, J., Cooley, R., Deshpande, M. and Tan, P. (2000) ‘Web usage mining: discovery and
applications of usage patterns from web data’, ACM SIGKDD Explorations Newsletter, Vol. 1,
No 2, pp.12–23.
Takács, O., Šarmanová, J. and Kostolányová, K. (2011) ‘Analysis of learning styles for adaptive
e-learning’, Digital Information Processing and Communications, Vol. 188, pp.368–376.
Thiyagarajan, C., Ravichandran, M., Suresh, R. and Kulanthaivel, G. (2011) ‘Effect of mining
educational data to improve adaptation of learning in e-learning system’, International
Conference on Sustainable Energy and Intelligent Systems (SEISCON), pp.922–927.
Truong, H.M. (2016) ‘Integrating learning styles and adaptive e-learning system: current
developments, problems and opportunities’, Computers in Human Behavior, February,
Vol. 55, pp.1185–1193.
Xu, B-C. and Zhang, Y-Y. (2014) ‘An improved gravitational search algorithm for dynamic neural
network identification’, International Journal of Automation and Computing, August, Vol. 11,
No. 4, pp.434–440.
Zakrzewska, D. (2008) ‘Using clustering technique for students’ grouping in intelligent e-learning
systems’, HCI and Usability for Education and Work, Vol. 5298, pp.403–410, Édité par
Springer, Heidelberg, Berlin.