Available online at www.sciencedirect.

com

ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2019) 000–000

ScienceDirect www.elsevier.com/locate/procedia

Procedia Computer Science 148 (2019) 87–96

Second International Conference on Intelligent Computing in Data Sciences (ICDS 2018)

Combining supervised and unsupervised machine learning

algorithms to predict the learners’ learning styles
Ouafae EL AISSAOUIa, Yasser EL ALAMI EL MADANIb, Lahcen OUGHDIRa and
Youssouf EL ALLIOUIc*
LSI, Sidi Mohamed Ben Abdellah Univesity, Fez, Morocco
a

b
ENSIAS, Mohammed V Univesity, Rabat, Morocco
c
LS3M, FPK, Hassan First University,B.P. : 145, 25000, Khouribga, Morocco

Abstract

The implementation of an efficient adaptive e-learning system requires the construction of an effective student model that represents
the student’s characteristics, among those characteristics, there is the learning style that refers to the way in which a student prefers
to learn. Knowing learning styles helps adaptive E-learning systems to improve the learning process by providing customized
materials to students. In this work, we have proposed an approach to identify the learning style automatically based on the existing
learners' behaviors and using web usage mining techniques and machine learning algorithms. The web usage mining techniques
were used to pre-process the log file extracted from the E-learning environment and capture the learners' sequences. The captured
learners' sequences were given as an input to the K-modes clustering algorithm to group them into 16 learning style combinations
based on the Felder and Silverman learning style model. Then the naive Bayes classifier was used to predict the learning style of a
student in real time. To perform our approach, we used a real dataset extracted from an e-learning system’s log file, and in order to
evaluate the performance of the used classifier, the confusion matrix method was used. The obtained results demonstrate that our
approach yields excellent results.
© 2019 The Authors. Published by Elsevier B.V.
© 2019 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open
Peer-review access
under article under
responsibility of the
the CC BY-NC-ND
scientific license(http://creativecommons.org/licenses/by-nc-nd/3.0/)
committee of the Second International Conference on Intelligent Computing in
Peer-review under responsibility
Data Sciences (ICDS 2018). of the scientific committee of the Second International Conference on Intelligent Computing in
Data Sciences (ICDS 2018).

* Corresponding author. Tel.: +212-067-774-7443; fax: +212-052-349-0354.

E-mail address: [email protected]

1877-0509© 2019 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review under responsibility of the scientific committee of the Second International Conference on Intelligent Computing in Data Sciences (ICDS 2018).

1877-0509 © 2019 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the Second International Conference on Intelligent Computing in
Data Sciences (ICDS 2018).
10.1016/j.procs.2019.01.012
88 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96
2 Author name / Procedia Computer Science 00 (2019) 000–000

Keywords:Web usage mining techniques; machine learning algorithms; K-modes clustering algorithm; Naive Bayes classifier; adaptive e-
learning systems; Learning style model;.

1. Introduction

Students have different ways to deal with learning materials and internalize information, these differences
determine learning styles. So, every student is characterized by a learning style that determines his/her preferred way
to see, process, perceive and understand information. Knowing the learning style helps the E-learning systems to
provide personalized contents to students that fit their requirements and enhance the learning process.
Many solutions have been proposed to identify students’ learning styles, the traditional one consists in asking
students to fill in a questionnaire, this solution has notable drawbacks. Firstly, filling in a questionnaire is a boring task
that consumes a lot of time. Secondly, students aren't always conscientious of their learning styles and the importance
of the further use of questionnaires, which can lead them to give arbitrary answers, therefore, the results obtained from
the questionnaires can be inaccurate and might not reflect the real learning styles of the students. Thirdly, the results
obtained from the questionnaires are static, while the learning styles can be changed during the learning process.
To overcome these limitations, many automatic approaches have been proposed, that aim to detect the students'
learning style automatically based on the behaviors of the students while they are interacting with the E-learning
system. Detecting learning styles automatically has many advantages over the traditional approaches, on one side, it
isn't necessary to waste time in using questionnaires since the information can be extracted from the students’
interaction with the system. On the other side, the Learning styles detected by the automatic approaches are dynamic
and can be changed according to the students' behaviors, while in the traditional approaches the LSs are static.
The automatic detection of learning styles requires the use of a LS model. A LS model aims to classify the students
according to the way they prefer to learn with. Many LSMs have been proposed in the literature such as [1][2][3].
.According to the previous researches [4][5], the FSLSM is the most appropriate to implement adaptive e-learning
systems .our approach is based on the FSLSM, section 4 presents the reasons why we have chosen it.
In this work, we have proposed an automatic approach to detect the LSs automatically based on the student’s
behaviors. The students’ behaviors were captured from the E-learning platform’s log file and then transformed to a set
of sequences where each sequence is composed of the learning objects accessed by a student during a session. The
learning objects were matched to FSLSM’s learning styles combinations. The students’ sequences and the matched
learning objects were taken as the input to the K-mode algorithm, this unsupervised algorithm was used to map the
sequences to the LSCs. the mapped sequences were used then as a training dataset to predict the learning style using
Bayes naive classifier.
The rest of this paper is organized as follows. Section 2 describes the learning style model and the algorithms used
in our approach. Section 3 introduces a literature review of related work. Section 4 describes the methodology of our
approach. The experiments and results are presented in section 5; Finally, Section 6 presents our conclusions and
future works.

2. Background

2.1. Learning styles

A learning style refers to the preferential way in which the student perceives, processes, understands and retains
information. Due to personality and environmental factors, each student has his own preferred ways of learning, for
example, when doing an experiment, some students can understand by following verbal instructions, while others have
to physically practice the experiment themselves. These differences in students' learning styles should be considered
by the educational systems to enhance the learning process.
Many definitions have been given to the term learning style in the literature. [6] Defined LS as the characteristic
strengths and preferences in the ways learners take in and process information. In [7], authors defined it as the
“composite of characteristic cognitive, affective, and physiological factors that serve as relatively stable indicators of
how a learner perceives, interacts with, and responds to the learning environment. According to[8], LSs refer to “a
 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96 89
Author name / Procedia Computer Science 00 (2019) 000–000 3

complex manner in which, and conditions under which, learners most efficiently and most effectively perceive,
process, store, and recall what they are attempting to learn.”.

2.2. The Felder and Silverman Learning style Model

A learning styles model enables the classification of learners according to the way they learn. Many learning styles
models exist in the literature such as:[9][10][2]. In this work, we have based on the Felder and Silverman model.
The FSLSM uses the notion of dimensions where each dimension contains two opposite categories, and each
student has a dominant preference in each dimension's category .the four dimensions of the FSLSM are :processing
(active/reflective), perception (sensing/intuitive), input (visual/verbal), understanding (sequential/global). A learning
style is defined by combining one category from each dimension.
Active learners tend to retain and understand information best by doing something with the learning material.
Reflective learners prefer to think about the learning material quietly first. Active learners also prefer to study in a
group, while the reflective learners prefer to work individually.
Sensing learners like courses that deal with the real world facts .They tend to be more practical by doing hands-on
(laboratory) work and they are more competent than intuitive at memorizing facts. Intuitive learners don't like
materials that contain a lot of memorization and routine calculations, and they prefer the ones involving abstractions
and mathematical formulations. They may be better than sensors at innovating and grasping new concept.
The visual learners prefer to see what they learn by using visual representations such as pictures, diagrams, and
charts. While the verbal learners like information that are explained with words; both written and spoken.
Sequential learners tend to go through the course step by step in a linear way, with each step followed logically by
the next one. Global learners tend to learn in large jumps, by accessing courses randomly without seeing connections.

2.3. K-modes clustering algorithm

The k-modes is a clustering algorithm that aims to group similar categorical objects into k clusters. This algorithm
is an extension of k-means clustering algorithm with three modifications: replacing the Euclidean distance function
with the simple matching dissimilarity measure. Instead of means, it uses modes, and instead of updating the centroids
it updates modes using a frequency based method [9].
The following steps illustrate how to cluster a categorical data set X into k clusters:
Step 1: Randomly Select k initial modes, one for each cluster.
Step 2: compute the distance from each object to each mode using the dissimilarity Measures, and then associate
each object to the cluster whose mode is the nearest to it. This association defines the first k clusters.
Step 3: Update the modes of the newly defined clusters from Step 2 using theorem (1) described below and then
retest the dissimilarity between the objects and the updated modes. If an object is found such that its nearest mode
belongs to another cluster rather than its current one, reallocate the object to that cluster and update the modes of both
clusters.
Step 4: repeat step 3 until no object has moved to another cluster after a full cycle test of the whole data set.
The dissimilarity Measures and the theorem (1) that are used in the previous steps; are described below:

2.3.1. Dissimilarity Measures

To compute the dissimilarity between two objects X and Y described respectively by m categorical attributes values
(𝑥𝑥1 , 𝑥𝑥2 ... 𝑥𝑥𝑚𝑚 ) and (𝑦𝑦1 , 𝑦𝑦2 , … , 𝑦𝑦𝑚𝑚 ), the two following functions can be used:

0, 𝑥𝑥𝑗𝑗 = 𝑦𝑦𝑗𝑗
𝑑𝑑(𝑋𝑋, 𝑌𝑌) = ∑𝑚𝑚
𝑗𝑗=1 𝛿𝛿(𝑥𝑥𝑗𝑗 , 𝑦𝑦𝑗𝑗 ) 𝑤𝑤ℎ𝑒𝑒𝑒𝑒𝑒𝑒 𝛿𝛿(𝑥𝑥𝑗𝑗 , 𝑦𝑦𝑗𝑗 ) = { (1)
1, 𝑥𝑥𝑗𝑗 ≠ 𝑦𝑦𝑗𝑗

(𝑛𝑛𝑥𝑥 +𝑛𝑛𝑦𝑦 )
𝑗𝑗 𝑗𝑗
𝑑𝑑𝑥𝑥 2 (𝑋𝑋, 𝑌𝑌) = ∑𝑚𝑚
𝑗𝑗=1 𝛿𝛿(𝑥𝑥𝑗𝑗 , 𝑦𝑦𝑗𝑗 ) (2)
𝑛𝑛𝑥𝑥 𝑛𝑛𝑦𝑦
𝑗𝑗 𝑗𝑗
90 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96
4 Author name / Procedia Computer Science 00 (2019) 000–000

Where 𝑛𝑛𝑥𝑥𝑖𝑖 and 𝑛𝑛𝑦𝑦𝑗𝑗 are respectively the numbers of objects that have the categories 𝑥𝑥𝑗𝑗 and 𝑦𝑦𝑗𝑗 for attribute 𝑗𝑗.

2.3.2. How to select a mode for a set

Let 𝑋𝑋 = {𝑋𝑋1 , 𝑋𝑋2 , … , 𝑋𝑋𝑛𝑛 } a mode of 𝑋𝑋 is a vector 𝑄𝑄 = [𝑞𝑞1 , 𝑞𝑞2 , … , 𝑞𝑞𝑚𝑚 ] that minimizes the following equation:

𝐷𝐷(𝑄𝑄, 𝑋𝑋) = ∑𝑛𝑛𝑖𝑖=1 𝑑𝑑(𝑋𝑋𝑖𝑖 , 𝑄𝑄) (3)

Where 𝐷𝐷 can be either equation (1) or equation(2).

Theorem 1: the function 𝐷𝐷(𝑄𝑄, 𝑋𝑋) is minimized if 𝑓𝑓𝑟𝑟 (𝐴𝐴𝑗𝑗 = 𝑞𝑞𝑗𝑗 \𝑿𝑿) ≥ 𝑓𝑓𝑟𝑟 (𝐴𝐴𝑗𝑗 = 𝐶𝐶𝑘𝑘,𝑗𝑗 \𝑿𝑿) for 𝑞𝑞𝑗𝑗 ≠ 𝐶𝐶𝑘𝑘,𝑗𝑗 for all 𝑗𝑗 =
𝑛𝑛𝐶𝐶
𝑘𝑘,𝑗𝑗
1, … , 𝑚𝑚 where 𝑓𝑓𝑟𝑟 (𝐴𝐴𝑗𝑗 = 𝐶𝐶𝑘𝑘,𝑗𝑗 \𝑿𝑿) = is the relative frequency of category 𝐶𝐶𝑘𝑘,𝑗𝑗 in 𝑋𝑋.
𝑛𝑛

2.4. Naïve Bayes

The Naive Bayes is a supervised classifier and it is an extension of the Bayes theorem with two simplifications:
The first simplification is to use the conditional independence assumption. That is, each attribute is conditionally
independent of every other attribute given a class label 𝐶𝐶𝑖𝑖 .See the following equation:

𝑃𝑃(𝑎𝑎1 ,𝑎𝑎2 ,…𝑎𝑎𝑚𝑚 \𝐶𝐶𝑖𝑖 ) 𝑃𝑃(𝑎𝑎1 \𝐶𝐶𝑖𝑖 )𝑃𝑃(𝑎𝑎2 \𝐶𝐶𝑖𝑖 )…𝑃𝑃(𝑎𝑎𝑚𝑚 \𝐶𝐶𝑖𝑖 ) ∏𝑚𝑚
𝑗𝑗=1 𝑃𝑃(𝑎𝑎𝑗𝑗 \𝐶𝐶𝑖𝑖 )
𝑃𝑃(𝐶𝐶𝑖𝑖 \𝐴𝐴) = 𝑃𝑃(𝐶𝐶𝑖𝑖 ) ∗ = 𝑃𝑃(𝐶𝐶𝑖𝑖 ) ∗ = 𝑃𝑃(𝐶𝐶𝑖𝑖 ) ∗ (4)
𝑃𝑃(𝑎𝑎1 ,𝑎𝑎2 ,𝑎𝑎3 ,…𝑎𝑎𝑚𝑚 ) 𝑃𝑃(𝑎𝑎1 ,𝑎𝑎2 ,𝑎𝑎3 ,…𝑎𝑎𝑚𝑚 ) 𝑃𝑃(𝑎𝑎1 ,𝑎𝑎2 ,𝑎𝑎3 ,…𝑎𝑎𝑚𝑚 )

The second simplification is to ignore the denominator 𝑃𝑃(𝑎𝑎1 , 𝑎𝑎2 , 𝑎𝑎3 , … 𝑎𝑎𝑚𝑚 ). Because 𝑃𝑃(𝑎𝑎1 , 𝑎𝑎2 , 𝑎𝑎3 , … 𝑎𝑎𝑚𝑚 )appears in
the denominator of 𝑷𝑷(𝑪𝑪𝒊𝒊 \𝑨𝑨) for all values of i, removing the denominator will have no impact on the relative
probability scores and will simplify calculations.After applying the two simplifications mentioned earlier, we obtain
the Naive Bayes classifier which consists in labeling the object A with the class label 𝐶𝐶𝑖𝑖 that maximizes the following
equation:

𝑃𝑃(𝐶𝐶𝑖𝑖 \𝐴𝐴)~ 𝑃𝑃(𝐶𝐶𝑖𝑖 ). ∏𝑚𝑚

𝑗𝑗=1 𝑃𝑃(𝑎𝑎𝑗𝑗 \𝐶𝐶𝑖𝑖 ) , 𝑖𝑖 = 1, 2, … , 𝑛𝑛 (5)

3. Related works

Enormous solutions have been proposed for automatic detection of learning styles. Many of those solutions are
based on data-driven approaches. The data-driven approaches use algorithms from the field of data mining and
machine learning to construct a model from the existing students' behaviors and their actual learning styles and then
this model is used to identify the learning style for a new student. According to [11]the Bayesian network classifier is
one of the most widely adopted classifiers to infer the learning style.
[12] Used a Bayesian network (BN) in order to identify students’ learning styles. They identified various behaviors
that may be relevant to detect learning styles in a given E-learning system. Then, a BN was trained with data from 50
students, using initial probabilities based on expert knowledge. The trained BN was then evaluated using 27 students.
For each student, the
BN provides a probability that the student has a particular learning style preference. As a result, the approach
obtained an overall precision of 58% for A/R, 77% for S/I a 63% for S/G (the V/V dimension was not considered).
[13] Addressed this problem by building a BN able to detect the difficulty that learners face during the exploration
process; and then providing specific assessment to guide and improve the learner’s exploration of the available
material.
The decision tree is a classification algorithm that is also frequently used in the field of automatic detection of
learning styles. [14] Proposed an approach to automatically detect the learners' learning styles from weblogs of the
students using the Data Mining technique and the Decision Trees classifier. Kolb’s learning style theory was
incorporated to understand e-learners’ learning styles on the web. [15] Proposed an approach to classify VARK
(Visual, Aural, Read/Write, Kinesthetic) learning styles of learners by using Decision Tree C4.5 algorithm. Data
 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96 91
Author name / Procedia Computer Science 00 (2019) 000–000 5

concerning learning styles of learners were collected via a questionnaire responded by 1,205 students. The collected
data were then classified using the Decision Tree C4.5 algorithm.
Neural networks are also commonly used in the automatic detection of learning styles,[16]proposed an approach
that uses neural networks to identify and track learners learning styles in order to ensure efficient recommendation of
resources. Their work was based on Felder-Silverman's dimensions.[17] Proposed the use of a feed-forward ANN (a
3-layer perceptron) with back propagation under a supervised learning model to identify learning styles. Ten behavior
patterns were used as inputs such as what kind of reading material did the student prefer, does the student revise their
answers on exams prior to submission and does the student ignore, post or read forums? As output, the neural network
produces three values, resenting the learning styles on three of the four learning style dimensions of the FSLSM.
As can be noticed, most of the proposed approaches used FSLSM's dimensions and considered that there are 8
learning styles where each one corresponds to a dimension's preference. In reality, there are sixteen learning style
combinations obtained by combining one preference from each dimension. In our proposed approach we have
considered sixteen LSC instead of 8.

4. Methodology

In this work we have proposed an approach that uses the web usages mining techniques and machine learning
algorithms to construct a model based on existing students behaviors (Fig. 1.). This model can be used to identify the
learning styles using input students behaviors data.
The students’ behaviors were captured from the web log of the E-learning platform. After preprocessing the log file
using the web usages mining techniques, the learners’ sequences were extracted to be used as an input to a clustering
algorithm. The learners’ sequences are composed of the set of LOs accessed by the learner during the learning process,
each sequence contains the sequence id, session id, learner id, and the set of learning objects accessed by the learner
in a session.
Learners’ Learning style
sequences combinations
combinations

K-mode
Algorithm

Cluster Cluster Cluster Cluster

01 02 … 16

New Naive Learning style

sequences Bayes combinations

Fig. 1. Our approach

The idea that we have come up with in this work, is to use students’ behaviors related to different types of learning
objects to determine learning styles. The learning objects were mapped to learning styles using FSLSM, and then a
clustering algorithm is used to group the learners’ sequences with respect to their learning styles based on the selected
Los .for example, if a sequence contains LOs such as: video, image, chart then this sequence will be classified as
having a visual preference. In this work, as Fig. 1 illustrates, two machines learning algorithms have been used:
92 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96
6 Author name / Procedia Computer Science 00 (2019) 000–000

• K-modes algorithm: is used to cluster the learners’ sequences into FSLSM’s learning styles combinations.
• Naïve Bayes: is used to predict the learning style combination for a new sequence.

The next subsections are organized as follows: first, we will represent how to match learning objects with learning
style combinations, then we will describe how the clustering and classification algorithms were used in our approach.

4.1. Matching learning objects to learning styles combinations

The automatic detection of learning styles requires the use of a LSM, in our approach we have based on the FSLSM
for many reasons. Firstly; this model uses the concept of dimensions and describes thoroughly the learning style
preferences. This description mentions the types of learning objects that can be included in each learning style
preference, and this is an interesting characteristic to our work, because knowing the learning style preferences of
learning objects helps us to determine the learning style of learners’ sequences. Secondly; according to our approach
a LS can be changed over time, and then updated using a classifier technique. The FSLSM provides this possibility
by considering LS as tendencies and students can act in a non-deterministic way as pointed by[18].
Thirdly , the FSLSM is the most used in adaptive e-learning systems and the most appropriate to implement them
as mentioned by [5].
The FSLSM is composed of 4 dimensions, where each dimension contains two opposite learning styles preferences,
and each learner has a dominant preference in each dimension. Thus, to identify the learner' learning style; we have
to determine a combination composed of one learning styles preferences from each dimension. As a result we will
obtain sixteen combinations:
Learning Styles Combinations (LSCs) = {(A,S,Vi,G), (A,S,Vi,Seq), (R,S,Vi,G), (A,S,Ve,Seq), (A,S,Ve,G),
(R,S,Ve,Seq), (R,S,Ve,G), (A,I,Ve,G), (A,I,Vi,Seq), (A,I,Vi,G), (R,I,Vi,Seq), (R,I,Vi,G), (R,S,Vi,Seq), (A,I,Ve,Seq),
(R,I,Ve,Seq), (R,I,Ve,G)}.
The above 16 learning style combinations have to be matched to learning objects, because identifying the learning
style combination of a sequence depends on knowing the learning styles of its learning objects. Based on the matching
table presented in our previous works [19] [20] [21]; we have obtained the following table, where the mapped learning
objects are considered as feature values of the K-modes clustering algorithm.

Table 1. Matching learning objects to LSCs

Practical Material
Announcements
Assignments

References

Sequential
Cluster ID

Examples

Topic list
Exercise

Cluster
Images
Videos

Forum

Charts
Demo

Email
PDFs
PPTs

meaning

C01 R,I,Ve,G yes yes no yes yes yes yes yes No no yes yes no no yes no
C02 A-I-Ve-G yes yes yes yes yes yes yes yes No no yes yes no no yes no
C03 R-S-Vl-G yes yes no yes yes yes yes yes yes yes no yes no no yes no
C04 A-S-Ve-G yes yes yes yes yes yes yes yes yes yes no yes no no yes no
C05 R-I-Vi-G yes yes no yes yes yes yes yes No no yes yes yes yes no no
C06 A-I-Vi-G yes yes yes yes yes yes no yes No no yes yes yes yes no no
C07 R-S-Vi-G yes yes no yes yes yes yes yes yes yes no yes yes yes no no
C08 A-S-Vi-G yes yes yes yes yes yes no yes yes yes no yes yes yes no no
C09 R-I-Vel-Seq yes yes no yes yes yes yes yes No no yes yes no no yes yes
C10 A-I-Ve-Seq yes yes yes yes yes yes yes yes No no yes yes no no yes yes
C11 R-S-Ve-Seq yes yes no yes yes yes yes yes yes yes no no no no yes yes
C12 A-S-Ve-Seq yes yes yes yes yes yes yes yes yes yes no no no no yes yes
 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96 93
Author name / Procedia Computer Science 00 (2019) 000–000 7

C13 R-I-Vi-Seq yes yes no yes yes yes yes yes No no yes yes yes yes no yes
C14 A-I-Vi-Seq yes yes yes yes yes yes no yes No no yes yes yes yes no yes
C15 R-S-Vi-Seq yes yes no yes yes yes yes yes Yes yes no no yes yes no yes
C16 A-Se-Vi-Seq yes yes yes yes yes yes no yes Yes yes no no yes yes no yes

4.2. The K-mode clustering algorithm

The first step of our approach aims to group the learners’ sequences into 16 clusters, and then labeling the resulted
clusters with the appropriate LSCs. We have used the k-modes clustering algorithm because of its ability to deal with
categorical objects, since the learners’ sequences which were considered as the input to that algorithm have categorical
attribute values.
After extracting the learners’ sequences from the log file using web usage mining techniques, we can use them as
an input to the K-modes by turning them into a matrix with M rows corresponding to M sequences and sixteen columns
to store the attribute values where the attributes of each sequence correspond to the sixteen Los presented in the
previous mapping table (Table 1).
Let 𝑆𝑆 = {𝑆𝑆1 , . . 𝑆𝑆𝑖𝑖 , … , 𝑆𝑆𝑛𝑛 } be a set of n sequences (categorical objects), each sequence 𝑆𝑆𝑖𝑖 is defined by 16 attribute
values (𝑠𝑠𝑖𝑖1 , . . 𝑠𝑠𝑖𝑖𝑖𝑖 , … , 𝑠𝑠𝑖𝑖16 ).
And let 𝐴𝐴1 , . . 𝐴𝐴𝑗𝑗 ,…, 𝐴𝐴16 be the 16 attributes describing the n sequences where:
𝑨𝑨1 = video, 𝑨𝑨2 = PPTs, 𝑨𝑨3 = demo, 𝑨𝑨4, = Exercise, 𝑨𝑨5 = Assignments, 𝑨𝑨6 = PDFs , 𝑨𝑨7 = Announcements,
𝑨𝑨8 = References, 𝑨𝑨9 = Examples, 𝑨𝑨10 = Practical Material, 𝑨𝑨11 = Forum, 𝑨𝑨12 = Topic list, 𝑨𝑨13 = Images,
𝑨𝑨14 = Charts, 𝑨𝑨15 = Email, 𝑨𝑨16 = Sequential.
For each sequence 𝑺𝑺𝑖𝑖 , Each attribute 𝑨𝑨𝐣𝐣 has two possible categorical attribute values: {yes or no, yes if the 𝑗𝑗𝑡𝑡ℎ
learning object exists in the sequence, No if it doesn’t exist.
Therefore, each sequence is presented as (𝑠𝑠𝑖𝑖1 , . . 𝑠𝑠𝑖𝑖𝑖𝑖 … 𝑠𝑠𝑖𝑖16 ) where 𝑠𝑠𝑖𝑖j = {𝑦𝑦𝑦𝑦𝑦𝑦 𝑜𝑜𝑜𝑜 𝑛𝑛𝑛𝑛}
To perform the k-modes algorithm we used the R software framework for statistical analysis and graphics, and in
order to write and execute the R code easily, we used the 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 as a graphical user interface. The dataset employed
in our approach was extracted from the E-learning platform’s log file† of Sup’ Management Group ‡. This dataset
records 1235 learners' sequences. The following table displays the results obtained after running the k-modes with
the R.

Table 2. Result of the K-modes algorithm

Cluster ID C01 C02 C03 C04 C05 C06 C07 C08 C09 C10 C11 C12 C13 C14 C15 C16
Number of
sequences in 58 73 147 49 82 56 187 59 35 32 119 50 46 96 37 109
each cluster
Total 1235

The obtained clusters were labeled with the LSCs based on the minimum distances between the clusters’ Modes
and the LSCs’ vectors presented in table I. the distances are computed using the dissimilarity measure described in
section 1 .

4.3. Naive Bayes Algorithm

After applying the K-modes algorithm and labeling the sequences with the LSCs, the labeled sequences were used as
a training dataset to train the classification algorithm, and then using it to predict the LSC for a new sequence.

†
http://www.supmanagement.ma/fc/login/index.php
‡
http://www.supmanagement.ma/fc/
94 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96
8 Author name / Procedia Computer Science 00 (2019) 000–000

In our work we have applied the Naive Bayes classifier for many reasons. First of all, because it’s one of the most
efficient machine leaning algorithms, it learns fast and predicts equally so, and it doesn’t require lots of storage. A
very important characteristic of Naïve Bayes for our work is that it is a probabilistic classification method, therefore,
in our approach we consider that the learner’s LSC is not deterministic and not stationary since it can be changed
during the learning process in an unexpected way, thus we can measure the LSC for a given learner after each iteration
using a probabilistic method.
Given a sequence 𝑆𝑆i defined with 16 attributes (𝐴𝐴1 , . . 𝐴𝐴𝑗𝑗 , … , 𝐴𝐴16 ) , as it is described in the previous subsection
above, each attribute 𝑨𝑨𝑗𝑗 has two possible attribute values 𝑎𝑎j = {𝑦𝑦𝑦𝑦𝑦𝑦 𝑜𝑜𝑜𝑜 𝑛𝑛𝑛𝑛}depending on the existence of the 𝑗𝑗𝑡𝑡ℎ in
the sequence.
Given a set of classified labels C = {𝐶𝐶1 , 𝑐𝑐2 , … 𝐶𝐶16 } where 𝐶𝐶i corresponds to one of the sixteen LSCs.
According to the Naive Bayes Classifier, for each new sequence 𝑆𝑆i , we will assign the classifier label 𝐶𝐶i that
maximizes the following equation:

𝑃𝑃(𝐶𝐶𝑖𝑖 ). ∏16
𝑗𝑗=1 𝑃𝑃(𝑎𝑎𝑗𝑗 \𝐶𝐶𝑖𝑖 ) , 𝑖𝑖 = 1, 2, … , 16 (6)

To predict the LSC for new sequences, we used the Byes naïve in R with the package e1071, and in order to
evaluate the performance of the used classifier, we did an experiment using the confusion matrix in R with caret. The
following section describes the experiment steps and the obtained results.

5. Experiment and results

5.1. Confusion metrics for classification problems

To evaluate the performance of the naive Bayes classifier we have used the confusion matrix method because of
its ability to deal with the multi-class Classification problems .the confusion matrix is a specific table layout that
enables for each class label to visualize the number of true positives, true negatives, false positives and the false
negatives. The following metrics can be computed based on the confusion matrix:

𝑇𝑇𝑇𝑇𝑖𝑖 Precision (Positive Predictive value) is the fraction of true

𝑃𝑃𝑖𝑖 = :
𝑇𝑇𝑇𝑇𝑖𝑖 + 𝐹𝐹𝐹𝐹𝑖𝑖 positive instances among the predicted positive instances.
𝑇𝑇𝑇𝑇𝑖𝑖 Negative Predictive Value (NPV) is the fraction of true negative
𝑁𝑁𝑁𝑁𝑁𝑁𝑖𝑖 = :
𝑇𝑇𝑇𝑇𝑖𝑖 + 𝐹𝐹𝐹𝐹𝑖𝑖 instances among the predicted negative instances.
𝑇𝑇𝑇𝑇𝑖𝑖 Recall (Sensitivity) is the proportion of positive instances that
𝑅𝑅𝑖𝑖 = :
𝑇𝑇𝑇𝑇𝑖𝑖 + 𝐹𝐹𝐹𝐹𝑖𝑖 are correctly classified as positive.
𝑇𝑇𝑁𝑁𝑖𝑖
: Specificity (True negative rate) is the proportion of negative
𝑆𝑆𝑖𝑖 = instances that are correctly classified as negative.
𝑇𝑇𝑇𝑇𝑖𝑖 + 𝐹𝐹𝐹𝐹𝑖𝑖
𝑇𝑇𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎𝑎𝑎 Accuracy, defines the rate at which a model has classified the
𝐴𝐴 = : records correctly.
𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑜𝑜𝑜𝑜 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐

5.2. Experiment

The validation set approach involves randomly dividing the available set of observations into two parts, training
set and testing set or hold-out set. The model is fitted on the training set, and the fitted model is used to predict the
responses for the observations in the testing set. In our experiment, the available set corresponds to the dataset obtained
after performing the K-modes.in this dataset all the sequences are labeled with LSCs, this dataset is split into two parts
training and testing with size of 70% and 30% of data respectively. The naive Bayes is fitted on the training set, and
then used to predict the responses for the observations in the testing set. The predicted and actual class labels in the
testing set are given as parameters to the confusion matrix function .the confusion matrix has been computed using
the R with the package caret, the following subsection describes the obtained results.
 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96 95
Author name / Procedia Computer Science 00 (2019) 000–000 9

5.3. Results and discussion

The more the validation metrics are high the more the classifier is good. The captured picture below displays the
results obtained after running the confusion matrix using R with the package caret. As can be noticed, all the validation
metrics: The Accuracy, the Recall (Sensitivity), the Specificity, the Precision (PPV) and the NPV have high values,
so we can say that the classifier used in our approach has well performed.

Fig. 2. The results obtained after applying the confusion matrix to the predicted and actual class labels in the testing set

6. Conclusion

Adaptive E-learning has become a promising solution to enhance the efficiency of online educational systems. A
necessary requirement in this solution is the automatic detection of learners' learning style in order to provide well-
adapted learning contents.in this context, we have proposed an automatic approach to identify the learners' learning
styles using web usage mining techniques and machine learning algorithms.
The E-learning platform’s log file was pre-processed using web usage mining technique in order to extract the
learners' sequences. Those sequences were mapped to learning styles combinations using the K-modes clustering
algorithm and based on the FSLSM. The labeled sequences were used as training set to train the naive Bayes classifier
and predict the learning style combination of a new student. To evaluate the performance of the used classifier we
have used the confusion matrix method. The produced results show that our approach performs well. In the future
work, we would like to have made a comparative study between the naive Bayes classifier and other machine learning
techniques such as the Bayesian network and decision tree.

References

[1] H. K. Laschinger and M. W. Boss, “Learning styles of nursing students and career choices.,” J. Adv. Nurs., vol. 9, no. 4, pp. 375–80, 1984.
[2] R. M. Felder and L. K. Silverman, “LEARNING AND TEACHING STYLES,” J. Eng. Educ., vol. 78, pp. 674–681, 1988.
[3] Biggs, “Study Process Questionnaire Manual. Student Approaches to Learning and Studying.,” Aust. Educ. Res. Dev., p. 53, 1987.
[4] S. Graf and Kinshuk, “Advanced adaptivity in learning management systems by considering learning styles,” in Proceedings - 2009
IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology - Workshops, WI-IAT Workshops 2009, 2009,
vol. 3, pp. 235–238.
[5] J. Kuljis and F. Liu, “A Comparison of Learning Style Theories on the Suitability for elearning,” Conference on Web Technologies,
96 Ouafae EL AISSAOUI et al. / Procedia Computer Science 148 (2019) 87–96
10 Author name / Procedia Computer Science 00 (2019) 000–000

Applications, and Services. pp. 191–197, 2005.

[6] R. M. Felder, “Matters of style,” ASEE Prism, vol. 6, no. 4, pp. 18–23, 1996.
[7] J. W. Keefe, “Learning style: An overview,” Student Learn. styles Diagnosing Prescr. programs, vol. 1, pp. 1–17, 1979.
[8] W. B. James and D. L. Gardner, “Learning styles: Implications for distance learning,” New Dir. Adult Contin. Educ., vol. 1995, no. 67, pp. 19–
31, 1995.
[9] Z. Huang, “A Fast Clustering Algorithm to Cluster Very Large Categorical Data Sets in Data Mining,” Res. Issues Data Min. Knowl. Discov.,
pp. 1–8, 1997.
[10] P. Honey and A. Mumford, The manual of learning styles, 3rd ed. Maidenhead, United Kingdom: Peter Honey Publications, 1992.
[11] J. Feldman, A. Monteserin, and A. Amandi, “Automatic detection of learning styles: state of the art,” Artif. Intell. Rev., vol. 44, no. 2, pp. 157–
186, 2015.
[12] P. García, A. Amandi, S. Schiaffino, and M. Campo, “Evaluating Bayesian networks’ precision for detecting students’ learning styles,”
Comput. Educ., vol. 49, no. 3, pp. 794–808, 2007.
[13] A. Bunt and C. Conati, “Probabilistic student modelling to improve exploratory behaviour,” User Model. User-Adapted Interact., vol. 13, no.
3, pp. 269–309, 2003.
[14] A. A. Kalhoro, S. Rajper, and G. A. Mallah, “Detection of E-Learners’ Learning Styles: An Automatic Approach using Decision Tree,” Int. J.
Comput. Sci. Inf. Secur., vol. 14, no. 8, pp. 420–425, 2016.
[15] O. Pantho, “Using Decision Tree C4 . 5 Algorithm to Predict VARK Learning Styles,” vol. 24, no. 2, pp. 58–63, 2016.
[16] B. Hmedna, A. El Mezouary, and O. Baz, “Identifying and tracking learning styles in MOOCs: A neural networks approach,” Eur. MENA
Coop. Adv. Inf. Commun. Technol., vol. 520, no. 2, pp. 125–134, 2017.
[17] J. E. Villaverde, D. Godoy, and A. Amandi, “Learning styles’ recognition in e-learning environments with feed-forward neural networks,” J.
Comput. Assist. Learn., vol. 22, no. 3, pp. 197–206, 2006.
[18] Kinshuk, T. C. Liu, and S. Graf, “Coping with mismatched courses: Students’ behaviour and performance in courses mismatched to their
learning styles,” Educ. Technol. Res. Dev., vol. 57, no. 6, pp. 739–752, 2009.
[19] Y. EL ALLIOUI, “Advanced Prediction of Learner’s Profile based on Felder Silverman Learning Styles using Web Usage Mining approach
and Fuzzy C-Means Algorithm,” Int. J. Comput. Aided Eng. Technol., vol. “in press,” no. forthcoming.
[20] O. El Aissaoui, Y. El Madani El Alami, L. Oughdir, and Y. El Allioui, “Integrating web usage mining for an automatic learner profile detection:
A learning styles-based approach,” in 2018 International Conference on Intelligent Systems and Computer Vision (ISCV), 2018, pp. 1–6.
[21] O. El Aissaoui, Y. El Alami El Madani, L. Oughdir, and Y. El Allioui, “A fuzzy classification approach for learning style prediction based on
web mining technique in e-learning environments,” Educ. Inf. Technol., vol. in press, pp. 1–17, Sep. 2018.