Expert Systems With Applications 174 (2021) 114774

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Learning style detection in E-learning systems using machine

learning techniques
Fareeha Rasheed *, Abdul Wahid
Department of Computer Science & Information Technology, Maulana Azad National Urdu University, Gachibowli, Hyderabad, India

A R T I C L E I N F O A B S T R A C T

Keywords: Learning style plays a vital role in helping students retain learned concepts for a longer time and also improves
Machine learning the understanding of the concepts. Learning styles in offline and online scenarios are recognized using ques­
Classification tionnaires. The recent trend is to identify and use attributes to detect the learning style of the learner auto­
Learning style
matically without disturbing the learner. The paper is an extension of the authors’ earlier work with some
E-learning
changes to the methodology. In this paper, the authors have identified new attributes and scaled-down the at­
tributes identified earlier, which would help identify the learner’s learning style. The authors implemented
classification algorithms and compared the accuracy of the different algorithms on the dataset. Various inter­
esting patterns are observed in learner’s behaviour while learning different types of concepts in different
situations.

1. Introduction consisted of four learning styles: Activists, Theorists, Pragmatists, and

Reflectors (Honey & Mumford, 1989). Learning modalities refer to how
Personalization has been a part of computing since the past few de­ learners use their senses to learn. The commonly considered modalities
cades, and each system currently being developed provides some sort of are four: visual (seeing), auditory (hearing), kinaesthetic (moving), and
personal touch to the users. E-learning systems have evolved a lot, from tactile (touching). Auditory learners enjoy reading or listening; visual
simple computerized instruction and quizzes to adaptive virtual envi­ learners often learn by seeing things. The kinaesthetic learn by doing
ronments (Wild et al., 2014). Learning styles are a category under a something rather than just observing, and the tactile learns better when
broader umbrella of learner characteristics; learner characteristics are they touch things and experience its structure and feel (Fleming, 2001).
the characteristics of the learner that affect the learning process Felder–Silverman’s learning style was initially researched by
(Nakayama & Santiago, 2012). Learning style is an interesting attribute Richard Felder and Linda Silverman to help professors in their teaching
in the learning scenario. Learning style is the preferred way of using methodology to engineering students. But now it has become a widely
one’s ability to learn (Sternberg, 1994). Different psychologists have accepted learning style model and is increasingly being adopted in many
given us different models of learning styles. Some of them are David online learning systems. The four dimensions defined in this learning
Kolb’s model, Peter Honey and Alan Mumford’s model, learning mo­ style model are perception toward information, input preferred, pro­
dalities, Neil Fleming’s VAK/VARK model, Anthony Gregorc’s model, cessing information, and understanding information. Perception has two
NASSP model, Felder-Silverman learning style, Gardner’s theory of variants sensing (concrete) and intuitive (abstract/imaginary). Inputs
multiple intelligence and so on (Pashler, McDaniel, Rohrer, & Bjork, can be of two types visual or verbal, and the information processing can
2008). is done in two ways. Processing of information happens in two ways
David Kolb’s model is also known as experiential learning. Kolb’s active (experimentation) and reflective (observing). Understanding in­
learning style defines different approaches to experiences. They are formation is also done in two ways sequential (follows a sequence) and
Concrete Experience and Abstract Conceptualization, as well as two global (does not work in any order) (Felder & Silverman, 1988). Fig. 1
related approaches toward transforming experience: Reflective Obser­ shows the various dimensions and their values in the Felder-Silverman
vation and Active Experimentation (Kolb & Kolb, 2005). Honey and learning style model.
Mumford’s learning style is based on Kolb’s learning style theory. It There can be 16 combinations of learning styles that can be formed

* Corresponding author.
E-mail addresses: [email protected] (F. Rasheed), [email protected] (A. Wahid).

https://doi.org/10.1016/j.eswa.2021.114774
Received 3 September 2020; Received in revised form 3 December 2020; Accepted 21 February 2021
Available online 25 February 2021
0957-4174/© 2021 Elsevier Ltd. All rights reserved.
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Fig. 1. Felder-Silverman learning style model, its dimensions and value of each dimension.

Fig. 2. Learning styles.

using values of each dimension in the model. These combinations can be and environmental dimensions which cannot be observed in online
formed by selecting one value from each of the dimension. This value in mode) were difficult to be mapped to online learning which makes it
questionnaires is attained by cross-checking the marking key to the difficult to completely avoid the usage of questionnaires. Hence, these
student’s answers in the questionnaire. For example, the cell 1 of the two models were selected Table 1 shows the learning styles according to
combinations grid in Fig. 2 asserts that a learner is an active learner the Gardner’s theory of Multiple Intelligences.
(participates in groups) is a sensing learner (focusses on facts), takes in Learning style is identified by using the questionnaires and the
visual input (learns well with images) and has sequential understanding scoring material given by these eminent psychologists. Learners fill the
(learns concepts in sequence rather than the complete picture). questionnaires, and they are marked. These later evolved into online
Gardner’s theory of multiple intelligences is a cognitive approach to questionnaires being marked automatically by computer programs
learning. This theory has emerged from cognitive research and “documents (Vincent & Ross, 2001). The advantage of questionnaires and online
the extent to which students possess different kinds of minds and therefore quizzes is that they are simple to implement and more straightforward to
learn, remember, perform, and understand in different ways” (Gardner & analyse. However, there are also problems associated with them. These
Hatch, 1989). questionnaires contain an average of more than twenty questions. There
Learning styles pertaining to each of the dominant intelligence in is a lot of bias when learners fill the questionnaires. Time constraints
learners expresses the best way each learner who has a particular
dominant intelligence learns. For example, a learner with mathematical-
Table 1
logical as the dominant intelligence learns best through numbers, logic
Gardner’s theory of multiple intelligences and learning styles.
and reasoning of concepts rather than words. There are other models
Dominant Learning style/learns best through
that identify the learner’s learning style using different aspects of
intelligence
learning. Each model has its own dimensions and values which can be
used to form combinations of learning styles. However, analysing all of Visual-spatial Learns best through images, pictures and diagrams
Verbal-linguistic Learns best through written or spoken word
them in a single study is difficult and out of the scope of this paper. Logical- Learns through numbers, quantities, logic and reasoning
Following are the simple reasons for selecting these two learning style mathematical
models—Firstly, most of the research done in the area worked with a Bodily- Learns best when does, acts, or works with simulations and
variety of learning style models, but very few of them worked with the kinaesthetic concrete experiences
Interpersonal Prefers to learn in solitude, dwell on concepts and theories
Gardner’s theory of multiple intelligences. Secondly, most of the work
Intrapersonal Prefers to learn in groups while communication, thinks out
done with the FSLM model skipped some of the dimensions (such as loud and learns while discussing
understanding dimension) and many researchers didn’t achieve higher Musical-rhythmic Learns through patterns, rhythm and listening
accuracies. Thirdly, some of the characteristics of students in other Naturalistic Learns best while making connections between concepts and
learning style models (For example, Dunn and Dunn has physiological environment, application-oriented learning
Existential Learns through a big-picture understanding of the concept

2
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

also pose a threat to the manual filling of questionnaires. Thus, the ac­ and generate simple if-then rules to detect the learning style. The evi­
curacy of the detected learning style may be less. dence of researchers using literature-based methods can be found in
Learning style is not dynamic; it does not change at regular intervals. (Carver, Howard, & Lane, 1999; Dung, & Florea, 2012; Graf, Liu, &
But there is always a better way of learning different things in different Kinshuk., 2010; Sabine Graf, 2006; Latham, Crockett, McLean, &
ways. It is better to analyse the learning style of the learner at regular Edmonds, 2012; Popescu, 2009; and Sangineto, Capuano, Gaeta, &
intervals, and the learners may not want to fill the long questionnaires Micarelli, 2008). The researchers using data-driven approaches have
again and again. There must be a method that uses intelligent techniques used different classification algorithms such as Bayesian networks, De­
to identify the learning style of the learner automatically. The significant cision Trees, Neural Networks, and other models. All these methods
challenges in the area are that; researchers have been using the item extracted some attributes from the user behaviour and built classifiers
response theory with questionnaires to identify the learning styles of the using the extracted data.
learners. Another problem identified during the review was that re­ Alkhuraiji, Cheetham, and Bamasak (2011), Carmona, Castillo, and
searchers overlook one or two dimensions of the Felder-Silverman Millán (2008), and García, Schiaffino, and Amandi (2008), García,
model of learning styles. A third problem points out that many Amandi, Schiaffino, and Campo (2007) are the researchers that have
learning style models are also not explored by the researchers. Re­ built Bayesian networks to detect learning styles. Decision Tree is a
searchers have been working in this area of automatically detecting the method that works in two phases, first building the tree and then
learning style of the learners for a long time. Some of the work is pruning the tree. A decision tree was used by the following researchers
reviewed and presented in the next section. The organization of this Özpolat and Akar (2009), Crockett, Latham, and Whitton (2017), and
paper is as follows: the next section, Section 2, shows the literature re­ Cha et al. (2006). The neural network is a machine learning technique
view, Section 3 presents the proposed methodology. Section 4 discusses that is based on the concept of neurons and the learning capability of the
the implementation and results; Section 5 discusses the conclusion and human brain. Neural networks have been used by the following re­
future work, which is followed by references. searchers to recognize learning style: Kolekar, Sanjeevi, and Bormane
(2010), Villaverde, Godoy, and Amandi (2006), Georgiou and Makry
2. Literature review (2004), and others. Some of the recent work which has been done in this
area is summarized in Table 2.
Before discussing the methods used to identify the learning style, a Table 2 below summarizes the methods and the shortcomings of the
quick account of what is learning style, thinking style, learning patterns, reviewed papers in the area of automatic detection of learning styles.
characteristics of learners with a particular learning style, whether they These papers have used different types of learning style theories. Most of
are malleable or not, and learning styles in online learning is needed. In the papers reviewed and summarized in Table 1 have used the Felder-
addition to the learning style models discussed in the previous section, Silverman model, and they are Ignacio, Rodr, Lugo, Castro, and Kono
there are a few researchers in the area whose work is worth discussing. (2015), Pitigala Liyanage, Gunawardena, and Hirakawa (2016),
Among them is Jan D. Vermunt, who proposed an integrated approach Hmedna, Mezouary, Baz, and Mammass (2016), Hmedna, Mezouary,
to learning style. The final version of the questionnaire he designed and Baz (2017), Racelis and Iii (2016), Bernard et al. (2017), Aissaoui,
consisted of 120 questions that cover the following learning compo­ Oughdir, and Allioui (2019), Sheeba and Krishnan (2019), Bursac,
nents: cognitive processing strategies, metacognitive regulation strate­ Milosevic, and Mitrović (2019) and Azzi and Jeghal (2019). Fatahi,
gies, conceptions of learning, and learning orientations(Vermunt & Moradi, and Farmad (2015) used the Meyers Brigg Test Indicator,
Vermetten, 2004). There are various scales and values in the question­ Rajper, Shaikh, Shaikh, and Mallah (2016) used the Kolb’s experiential
naire, in the first component, we have deep processing and step-wise learning model and Hasibuan, Nugroho, and Santosa (2019) used the
processing, and in turn, each of them has different extremes. The VAK model.
regulation strategies have two scales self-regulation and external regu­ Since the prediction and classification involve machine learning
lation with their values. The questionnaire helps us gain a complete models and algorithms, a short account of machine learning and its
insight into a learner’s mind; it has high validity and reliability and has models and details of models are mentioned in this and the following
been widely used. However, a questionnaire of this kind, and depth, paragraph. Machine learning is a convergence of statistics and computer
seems difficult to understand in an online learning context. Few of its science wherein machines learn to improve performance from their
concepts, such as external regulation and learning orientations even previous experience, a manner similar to humans. The only difference is
when monitored in online learning, are challenging to include in online that computers learn from data, and humans learn from situations and
learning, and it would use up resources. Extensive research must be done experiences (Bishop, 2006). Machine learning consists of 3 types of
to see how this can be incorporated into online learning. learning algorithms, supervised, unsupervised, and reinforcement
Learning styles and their malleability has been researched and dis­ learning. Supervised learning happens when data along with labels are
cussed by Zhang (2011), where the researcher has discussed how supplied to the computer as input; the machine learns patterns from it
learning style is affected by culture. Cognitive styles are also malleable, and tries to predict the label of newly provided data. For example,
as proven by Angeli and group, using an experiment with children aged depending on the temperature, humidity, precipitation, and other
6 to 7 years (Angeli, Valanides, Polemitou, & Fraggoulidou, 2016). weather factors, the computer will predict whether it will rain today or
Researchers have been using different methods and have based their not after learning from the data. Unsupervised learning doesn’t come
research on different learning style models. Some of the approaches to with any labels; the machine has to study the patterns in the data and
identify learning style and the frequency of usage of learning style group them into groups according to the characteristics of the data. For
models by various researchers are summarized in this section. Re­ example, tagging of tweets automatically under different types of sub­
searchers have results that prove the link between learning style and jects. Reinforcement learning works on the concept of reward theory;
academic performance (Lynch, Woelfl, Steele, & Hanssen, 1998; from a particular state, a machine moves forward if the solution is
Komarraju, Karau, Schmeck, & Avdic, 2011; Cassidy & Eachus, 2000). correct or goes back to the previous state if the solution is wrong
Researchers have proposed many artificial intelligence techniques to (Marsland, 2015).
detect the learning style of the learners automatically. There are two Classification is a technique wherein we divide the data into classes
different approaches defined by Bernard, Chang, Popescu, and Graf and give a label to it. There are many algorithms in machine learning to
(2017) literature-based and data-driven approach for automatic detec­ perform classification. The basic ones are Naïve Bayes, Support Vector
tion of learning style. Data-driven strategies aim at building classifiers Machine, K-Nearest Neighbours classifier, Decision Trees, and Logistic
based on the data. Regression. Then there are ensemble algorithms such as random forest,
Literature-based approaches use the user model to get hints from it Ada Boost, and Bagged Classifier (Kotsiantis, Zaharakis, & Pintelas,

3
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Table 2
A Summary of the Current Work in the Area of Automatic Learning Style Detection.
Paper Method Used Description

Framework for Automatic Identification of Learning Styles in Learning Fuzzy classification The results obtained from the classification are not compared with precision and
Management Systems, 2015, (Ignacio et al., 2015) accuracy
Behavioural Feature Extraction to Determine Learning Styles in E- K-means clustering The research aimed to use student behaviour in online learning to study their
Learning Environments, 2015, (Fatahi et al., 2015) learning pattern. The research was done in a blended learning model. However,
the results are not compared to different approaches.
Detecting Learning styles in Learning Management Systems Using Data Decision Tree J48 classifier performed better than many other approaches, and the proposed
mining, 2016, (Pitigala Liyanage et al., 2016) method performs better than the existing approaches.
Automatic Detection of Learning Styles on Learning Management Bayesian Network The accuracy of results for all the dimensions of the Kolb’s inventory is around
Systems using Data Mining Technique, 2016, (Rajper et al., 2016) 71%; the results can be improved and compared with other classifiers
A Machine Learning Approach to Identify and Track Learning Styles in Neural Network The authors have implemented a neural network to classify the learners in
MOOCs, 2016, (Hmedna et al., 2016) different dimensions of the FSLM model, but results have not been compared.
Classification of Learning Styles in Virtual Learning Environment using Naïve Bayes, J48, The authors implemented four different algorithms to classify learners into
Data Mining: A Basis for Adaptive Course Design, 2016, (Racelis & Iii, NBTree, BayesNet different classes of different learning styles. They also compared the accuracy of
2016) these different algorithms. They concluded that the J48 algorithm performed
better
Identifying and tracking learning styles in MOOCs: A neural networks Neural Network The method was not implemented, and the results were not compared.
approach, 2017, (Hmedna et al., 2017)
Learning Style Identifier: Improving the Precision of Learning Style Neural Network The authors used four different algorithms to identify the learning style of the
Identification Through Computational Intelligence Algorithms, 2017, learners, the proposed approach is compared to existing methods, and there was
(Bernard et al., 2017) an improvement in the accuracy.
Model Detecting Learning Styles with Artificial Neural Network, 2018, Neural Network The authors used Latent semantic indexing to use prior knowledge of the learner
(Hasibuan et al., 2019) to predict the learner’s learning style and then compared it to the data received
from the learners
A fuzzy classification approach for learning style prediction based on Fuzzy C means The authors, unlike previous researchers, considered all the sixteen learning style
web mining technique in e-learning environments, 2018, (Aissaoui categories, including all the dimensions of the FSLM model. The results of the
et al., 2019) classification are compared with results from K-means classification. It would be
interesting to observe the difference in results in other classifiers
Automatic Detection of Students Learning Style in Learning Decision Tree The overall accuracy of the model proposed is 87%. The results were not
Management System, 2019, (Sheeba & Krishnan, 2019) compared, and two dimensions of the model were not considered in modeling
Proposed Model for Automatic Learning Style Detecting Based on Fuzzy C means and The proposed model provides exceptional accuracy; however, the model is not
Artificial Intelligence, 2019, (Bursac et al., 2019) neural networks validated using real data sets
A robust classification to predict learning styles in adaptive E-learning Fuzzy C means The accuracy of the technique is 93% approximately; however, the results have
systems, 2019, (Azzi & Jeghal, 2019) classifier not been compared to other classification techniques.

2007). There are approaches to using artificial neural networks for the machine has already learned from data. There are a lot of applica­
classification, and they outperform every other machine learning algo­ tions of classifications, serious ones such as disease prediction, traffic
rithm. There are problems associated with it. Some of the issues are the monitoring, and simpler ones such as spam email detection, and games
ad hoc nature of the neural networks, the requirement of a large dataset, (Bernadó-Mansilla & Garrell-Guiu, 2003). Naive Bayes classifiers are a
missing validation procedures. Other challenges are getting the perfect collection of classification algorithms based on Bayes’ Theorem. The
weights; time for convergence is large even when we solve simple algorithms share a universal principle that every pair of features being
problems; a lot of hardware is required (Vemuri, 1993). Although a lot of classified is independent of each other (Rish, 2001). A Support Vector
issues are already solved, some of these exist, and neural networks are a Machine (SVM) is a supervised machine learning algorithm that can be
tough choice when data is less. employed for both classification and regression purposes. SVMs are
Classification in machine learning is used to make predictions after based on the idea of finding a hyperplane that best divides a dataset

Table 3
Attributes and Description Used to Deduce the Dominant Intelligence.
Dimension Attribute name Description

Verbal-Linguistic T_a Time spent on Audio files

H_wiki Hits to the wiki material for the related resources
T_podcast Time spent on podcast
T_read Time spent on reading material

Mathematical-Logical T_conceptmaps Time spent on studying concept maps

T_charts&Graphs Time spent on studying charts and graphs
T_abstractconcepts Time spent on studying abstract concepts
N_Seq_theory_before_examples Number of times theory was studied before examples

Visual-Spatial T_image Time spent on studying images

N_videoconferences Number of conferences joined w.r.t total conferences proposed
T_ppt Time spent on study using PowerPoint presentations
N_youtubevideos Number of YouTube videos accessed related to the content

Bodily-Kinesthetic T_simulation Time spent in studying simulations

N_applicationlevelquestions Number of questions of the application level correctly solved
T_casestudies Time spent on studying case studies
T_interactiveobjects Time spent on using interactive objects

Interpersonal-Intrapersonal T_forum_reading Time spent in forum reading posts

N_posts Number of posts written or answered in the discussion forum
N_groupdiscussions Number of group discussions joined
T_selfassess Time spent on self-assessment exercises

4
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Table 4
Attributes and their Description Used to Identify the Learning Style According to FSLM Model.
Dimension Attributes Description

Input Visual/Verbal T_Image Time spent on study using images

T_video Time spent on watching video content
T_read Time spent on reading material
T_audio Time spent on listening to audio content
N_exercise_after_read Number of exercises completed after study using reading material
N_exercise_after_graphic Number of exercises completed after study using graphics
T_hierarchies Time spent on studying hierarchies
T_powerpoint Time spent on study using PowerPoint presentations
∑[ ]
Perception Sensing/Intuitive T_abstract Time spent on abstract content calculated as:Tabstract = Tfacts +Tcasestudies +TTheory
∑[ ]
T_concrete Time spent on concrete content calculated as:Tconcrete = Texamples +Tcasestudies +Tcomplexconcepts
T_result Time spent on viewing and understanding result
N_standard_questions_correct Number of standard questions answered correctly
N_creative_questions_correct Number of creative questions answered correctly

Processing Active/Reflective N_msgs_posted Number of messages posted in the discussion forum

N_exercises_visited Number of exercises visited by the learner
T_reading_in_forum Time spent in reading discussions in the forum
T_solve_exercise Time taken to solve exercise (high in case of reflective and low in case of active)
T_submit_assignment Time taken to submit an assignment
N_groupdiscussions Number of group discussions joined
T_outlines Time spent in study using outlines

Understanding Sequential/global Skipped_Los Whether learning objects were skipped by the learner or not. Binary value: yes or no
N_next_button_used Number of times the next button which is high for sequential learners and low for global learners
T_spent_in_session Time spent in a session; it is high for global learners as they learn in large leaps
N_questions_on_details Number of questions attempted that deal with details of a concept
N_questions_on_outlines Number of questions that deal with the outlines of concepts

-when plotted on the graph into two classes (Hsu & Chang, 2003). K- learning styles. The study by Jegatha and group (Jegatha Deborah,
nearest neighbours (KNN) algorithm uses ’feature similarity’ to predict Baskaran, & Kannan, 2014), helped us to understand the link between
the values of new data points, which means that the new data point will the theory of learning styles and e-learning. Other interesting studies in
be assigned a value based on how closely it matches the data points in Özpolat & Akar (2009), Khan, Shamim, & Nambobi (2018), and Santo
the training set (Sutton, 2012). Decision Trees are flowchart like struc­ (2006) discussed common behaviour, learning preferences, recall and
tures, for predicting a class label for a data point, we start from the root retention rates, efficiency and performance of learners with various
of the tree. We compare the values of the root attribute with the data learning styles. A study (Huang, Lin, & Huang, 2012) analysed the
point’s attribute. Based on the comparison, we follow the branch cor­ performance of learners when they learnt using resources according to
responding to that value and jump to the next node (Kearns & Mansour, their learning style. The study is promising as it proves that presenting
1999). Each algorithm has parameters that need to be fine-tuned to learners with customized resources to learn betters their performance.
achieve optimal results of accuracy, precision, recall, and other metrics There is a study on linking multiple intelligences and online learning
during the testing and the training phase. activities (Zhang & Bonk, 2009) helped us identify what learners with
various learning styles preferred to do in an online learning system.
2.1. Problems identified Students’ typical choice of resources, the time taken to learn each
concept using their choice, their performance in assessments and other
The authors, in their previous work in (Rasheed & Wahid, 2019), learner attributes were observed. We then extracted indicators from
proposed a neural network for the identification of learning styles based learner’s online behaviour that will help us to predict their learning
on multiple intelligences. The proposed method was not validated, and style. After the extraction of learner behaviour from learner logs, we
when it was validated, it generated inaccurate models. Hence, to over­ preprocessed it. The processed data was then passed as input to the
come the problems that occurred from the previous research, this work machine learning classifiers. The Recognition system is based on ma­
has been proposed, trained, and validated. A lot of work has been done chine learning classifiers which predict the learning style of the learner
in the area of learning style detection; the below problems are pre­ based on the inputs provided. The indicators and attributes are
dominant in the area. mentioned in Tables 3 and 4. The extraction of indicators is necessary
because we want to eliminate the use of questionnaires to identify the
• Real datasets have not been used for the validation of the proposed learning style. To prove the elimination of the use of questionnaires, we
models. evaluate the consistency of questionnaire-based identification and the
• In some publications, the models have not been compared to others predicted learning style using the proposed attributes and trained
for accuracy, precision, and recall model. The predicted learning style can be used to personalize the
• Most of the papers that used the FSLM learning style model mostly learning resources and recommend them specific resources that cater to
skipped either one or two dimensions of the model their learning styles so that learning is optimal.
• Most of the publications we reviewed worked with the Kolb’s The proposed methodology is an extension of our work done in
learning style inventory and FSLM model; a very few of them used Rasheed and Wahid (2019). In the paper, we proposed attributes to
Gardner’s theory of multiple intelligences. recognize the type of intelligence based on the theory of Gardner’s
theory of multiple intelligences. We proposed a neural network archi­
tecture to recognize the intelligence, and the intelligences was linked to
3. Methods
learning styles. In this paper, we extend the work. Our earlier article
didn’t implement the proposed work; when we wanted to extend and
The study initially focussed on studying learning style models, their
peruse the work using deep learning, we were not able to achieve the
dimensions, values, and combinations of values to identify the learner’s

5
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Fig. 3. Research Design.

desired accuracy. Deep learning works best with a large dataset, our instance, learner ’A’ filled the questionnaire initially. A’s behaviour is
dataset was small and the trained network overfitted. To avoid this tracked in the online learning system. The indicators to extract the
overfitting and reduce the use of computational power, we preferred not proposed attributes are taken from log files. The trained model then uses
to use deep learning, and hence we made feature selection and removed the indicators to predict A’s learning style. Finally, to evaluate the
attributes that had a negative correlation with the output class. With a consistency of the proposed method, the results of the questionnaire and
more in-depth study of literature and the earlier identified attributes, we the predicted value is equated. If they are equal, the model is consistent,
then shortlisted some attributes. After removing some attributes from if not, the model is not consistent. The percentage of consistency for each
our previous work, we add new attributes. We also observe attributes dimension is discussed in the discussion section. The data and the study
that help to identify a learner’s behaviour in a discussion forum. The do not aim to compare the learning styles of learners from different
research proposed in this paper follows the research design shown in classes. The study rather focusses on developing a machine learning
Fig. 3. This module is a part of a more extensive system that performs the model to predict the learning style of learners. The data is then cleaned,
sequencing of learning resources. pre-processed; machine learning models are trained and tested to make
The research follows an experimental research design; the contrib­ predictions. The predictor predicts the learning style of the learner,
utors of data are the usual behaviour of the learner with a specific which is the first outcome. The outcome of exciting patterns in learner
learning style and student behaviour attributes in the E-learning system. behaviour with respect to learning style is achieved by carefully
Data is collected both through literature review and through online logs observing the learner logs in different situations.
of learners in the system. The data collected from learners of various The research methodology followed to predict the learning style is
classes intended to check whether our trained model can correctly presented in Fig. 4, which summarizes the steps taken to detect the
predict their learning style using the proposed attributes or not. For learning style of the learner. The actions performed are typical to any

6
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Fig. 4. Proposed Methodology.

data science-oriented machine learning technique used for classification We have used the relative time in comparison to the total files desig­
and prediction. nated for particular content. Values are coded. If a learner spends 75% or
The learners were presented with questionnaires before they began a more of the appointed time, value is 3. If the time spent is between 50%
course, on condition of anonymity, about two hundred students of and 75% value is 2, and below 50% value is 1. Table 3 contains the
various classes in a university participated in the experiment. They then attributes used for the deduction of the learner’s dominant intelligence
started a course on Moodle, which was modified to observe the multiple defined in Gardner’s theory of multiple intelligence. The common
attributes. The learner activity is monitored from the log facility in characteristic we used in different intelligence deduction is the learning
∑ ∑
Moodle. The unprocessed data was collected, and processing was done gain which is defined as Learninggain = ( PostTestscore − PreTestscore )/
∑
to clean the data. The processed data was then used to train and test the PreTestscore ((sum of pre-test score subtracted from the sum of post-test
classifiers. The classifiers we implemented were decision trees, Support score) divided by the sum of pre-test score)
Vector Machines, K-Nearest Neighbour, Naïve Bayes, Linear Discrimi­ Table 4 consists of the attributes used for the FSLM model; attributes
nant Analysis, Random Forest, and Logistic Regression. are repeatedly used because the FSLM model’s input dimension consists
These algorithms are commonly used for classification in machine of visual and verbal types.
learning; when the models are trained with enough data, they can be The inputs to the classification algorithm are the attributes in the
used for prediction. A short account of these algorithms and their normalized form; the data was observed from the learner logs, and
working is discussed in the literature review section. Using some attri­ values were normalized as discussed earlier. The output classes being
butes from our previous work and proposing new attributes, the the dominant value of intelligence. The attributes have been crafted
following attributes were identified. The attributes in Table 3 and 4 may after studying a lot of research in the area, a study of media associated
not be enough for this research, but they were selected because they can with each type of intelligence can be found in Kolås and Staupe (2007).
be either easily observed or derived from the learner logs. Making the Another example of designing a rule-base for e-learning recommenda­
attributes complex will create issues in the data collection phase, and a tions using multiple intelligences is developed in Kaewkiriya, Utakrit,
lot of coding should be done to make the interface observe more attri­ and Tiantong (2016). Igbrue and Pathak (2008) created e-learning
butes. We try to analyse the learner’s learning style according to two content using the Multiple intelligences theory with an attempt to in­
different learning styles theories. The first theory is the theory of mul­ crease the learning efficiency of learners. Authors in Mankad, Sajja, and
tiple intelligences, and the next one is the Felder-Silverman model. Both Akerkar (2011) used fuzzy logic and evolutionary rules to build a hybrid
these models are deemed efficient by sufficient proof in the literature to system to detect the multiple intelligence of the learner. The rules or
provide personalization in E-learning. parameters have been identified using the characteristics of learners
Table 3 and 4 describe the attributes that were considered for the with each dominant intelligence, the media selection, the time they
deduction of learning styles. We have made the following assumptions: spent on each media, and others.

7
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

The FSLM model is a 11 points scale. If the score on a scale is 1–3, 5. Discussions
then the user has mild preference, 5–7 is moderate preference and 9–11
is strong preference. The studies discussed in the literature review sec­ To discuss the results, we need to first define the metrics we have
tion give us a picture of what learners with particular learning style do in used for evaluating the models and the proposed approach as compared
online learning. If all/ most of the activities of the learner point towards to the self-report inventory of learning styles. Accuracy of a machine
the visual learning style, the learning style as predicted by the model is learning algorithm is the number of correct predictions divided by total
visual, and if all/ most of the activities of the learner point towards predictions made; accuracy is specific to machine learning models. The
verbal learning style, the learning style as predicted by the model is number of positive predictions divided by the total number of positive
verbal. This behaviour of machine learning models is common; machine class values that were predicted is precision. The recall is the number of
learning models help in predicting the class of a learner according to the positive predictions divided by the total number of positive class values
inputs supplied. However, these classifiers do not indicate the tendency in the test data. F1-score is the weighted average of precision and recall.
or intensity of any class. If both the activities/ actions are equal then The cross-validation score is used to check for the overfitting and
there is marking as visual-verbal. However, some ambiguities occur underfitting in algorithms. Accuracy is a term that is specific to machine
which will be resolved during our future work. learning models. To compare the results of the proposed recognition
We collected data from graduate students studying a bachelor’s de­ system and the self-report inventory we used another metric. The metric
gree in computer science. Metrics measures the quality of prediction chosen here is consistency. Consistency is used here as a measure of
performed by the machine learning algorithms. The most common evaluating how correctly our proposed model predicts the learning style
metrics are accuracy, precision, recall, F1-score, cross-validation score, of the learner using the attributes. % Consistency ((correct predictions
and area under the curve. There are metrics other than the mentioned by machine/No of students from manual marking) *100). For example,
above, but researchers use these while performing classification and if there are 45 students of the visual type identified in the manual
prediction. The area under curve and cross-validation also explain to us marking, we check out of the 45 how many were correctly labelled as
whether the results are significant or not. For training, we chose 70% of visual type using our proposed model.
the data; testing was done on 30% of data. For validation, we performed From Table 5, we can observe that, for the data collected from
cross-validation, and its results are present in Table 6 and 5. Gardner’s theory of multiple intelligences, the algorithm that performs
the best is the support vector machine with an accuracy of 75.55% and
4. Results with close values of precision and recall. The cross-validation score is
74%. The cv score proves that irrespective of the 70–30 division of data
For the attributes identified for the theory of multiple intelligences, or cross-validation, the accuracy of the algorithm will be around 76%,
we collected a dataset with 498 samples from students taking online which is acceptable. Therefore, in the final prediction of dominant in­
courses. The pre-processing was done. Then various classification al­ telligence in a real-time scenario, we used the support vector machines.
gorithms such as decision trees, Support Vector Machines, K-Nearest An exciting future direction would be conducting a pilot study and then
Neighbour, Naïve Bayes, Linear Discriminant Analysis, Random Forest, identify and reattribute characteristics for each dominant intelligence.
and Logistic Regression were applied. We run the classifiers on the py­ We can observe from the Table 6 that, for the input dimension, the
thon kernel 3.1. To avoid overfitting (A problem where the trained Support Vector Machine classifier has the highest accuracy of 85.22%.
technique predicts everything you enter into a particular class), we For the processing dimension, the Naïve Bayes performs better than the
tuned the parameters of the classifiers using the GridSearch technique of other algorithms. For the perception dimension, the Decision Tree
hyperparameter tuning. The GridSearch technique helped us gain an Classifier performs better with 92% accuracy. For the understanding
understanding as to how the training and testing accuracy differs with dimension, the results have improved when compared to Graf et al.
automatic tuning. We then performed manual tuning of the parameters (2010) and the highest accuracy has been achieved in the case of the K-
to achieve optimal results of accuracy, precision, and recall. The values Nearest Neighbor Classifier.
of the trained parameters can be seen in Table 5. The cross-validation score of each dimension and each algorithm is
For the attributes identified for the Felder Silverman Learning style in Table 7. It can be observed that even though the precision, recall, and
model, we collected a dataset with 498 samples from students taking accuracy values in size are the best for the support vector machine, the
online courses. The pre-processing was done, fit, and transform; the data cross-validation score is only 78%. Other algorithms which have high
were normalized between 0 and 1. We used the decision trees, Support cross-validation score are random forest, logistic regression, and linear
Vector Machines, K-Nearest Neighbour, Naïve Bayes, Linear Discrimi­ discriminant analysis. In real-time detection in the input dimension, we
nant Analysis, Random Forest, and Logistic Regression and to avoid preferred to use SVM. The reasons for doing so are that logistic regres­
overfitting; we performed tuning of the parameters. The algorithms sion and linear discriminant analysis are prone to overfitting. Unlike
were run on Python Kernel 3.1. Table 6 summarizes the accuracy of the training data, real-time data might have some relationships that are
various algorithms for different dimensions of the FSLM model. complex and are not needed. For the processing dimension, Naïve Bayes
performs the best concerning the accuracy, precision, and recall, but the
cross-validation score says it all; it is far less than SVM, which performs
third best when compared to Naïve Bayes. We preferred SVM for the
real-time prediction as the cross-validation score of SVM is high when
Table 5 compared to all other algorithms. For the understanding dimension, we
Summary of Results for the Dataset Collected with Attributes Identified for the
again preferred the SVM for the real-time prediction because its cross-
Theory of Multiple Intelligences.
validation score is the highest, and real-time data need not be similar
Algorithms Accuracy Precision Recall CV F1- to the datapoint from the testing or training datasets. For the perception
Score score
dimensions, we used the decision tree, although improved accuracy with
Support Vector Machine 75.55% 73% 76% 74% 73% the decision tree means that the classifier has learned complex re­
Decision Tree 45.55% 41% 46% 39% 43% lationships. But when the cross-validation scores are observed, they are
Logistic Regression 73.33% 71% 73% 74% 72%
Random Forest 73.33% 74% 73% 68% 73%
still high. Therefore, the decision tree classifier is used for real-time
K-Nearest Neighbours 67.77% 68% 68% 60% 67% prediction.
Linear Discriminant 69.44% 69% 69% 74% 69% How will this benefit the practitioners and researchers? This question
Analysis can be answered by discussing two crucial factors of this research. One,
Naïve Bayes 70.55% 65% 71% 71% 65%
the difference and similarity between human marking and machine

8
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Table 6
Summary of Results for the Dataset Collected with Attributes Identified for the FSLM Learning Styles Model.
Input Dimension Processing Dimension

Algorithms Accuracy Precision Recall F1-score Algorithms Accuracy Precision Recall F1-score

Support Vector Machine 85.22% 85% 85% 85% Support Vector Machine 85.33% 86% 85% 85%
Decision Tree 80.13% 80% 80% 80% Decision Tree 83.33% 85% 83% 84%
Logistic Regression 84.65% 84% 84% 84% Logistic Regression 85.33% 86% 85% 85%
Random Forest 80.13% 81% 80% 80% Random Forest 82.66% 83% 83% 83%
K-Nearest Neighbours 76.15% 76% 76% 76% K-Nearest Neighbours 82.66% 83% 83% 83%
Linear Discriminant Analysis 83.44% 83% 83% 83% Linear Discriminant Analysis 86.00% 86% 86% 86%
Naïve Bayes 77.48% 83% 77% 76% Naïve Bayes 86.66% 88% 87% 86%

Understanding Dimension Perception Dimension

Algorithms Accuracy Precision Recall F1-score Algorithms Accuracy Precision Recall F1-score
Support Vector Machine 83.33% 83% 83% 83% Support Vector Machine 91.33% 92% 91% 91%
Decision Tree 78.00% 79% 78% 77% Decision Tree 92.00% 92% 92% 92%
Logistic Regression 82.00% 82% 82% 82% Logistic Regression 91.33% 92% 91% 91%
Random Forest 82.00% 82% 82% 82% Random Forest 90.66% 91% 91% 91%
K-Nearest Neighbours 84.66% 85% 85% 85% K-Nearest Neighbours 88.00% 89% 88% 88%
Linear Discriminant Analysis 81.33% 81% 81% 81% Linear Discriminant Analysis 90.66% 92% 91% 91%
Naïve Bayes 79.33% 80% 79% 79% Naïve Bayes 87.33% 87% 87% 87%

Table 7 Table 8
Cross-Validation scores for each dimension and the algorithms. % Consistency between manual marking and machine prediction for Gardner’s
Algorithms Dimensions
theory of multiple intelligences.
Dominant Number of students Number of % Consistency
Input Processing Understanding Perception
Intelligence with intelligence students with ((correct
SVM 78% 83% 83% 89% from Manual Dominant predictions by
Random Forest 79% 77% 77% 91% Marking of E- Intelligence machine/No of
Decision tree classifier 73% 74% 74% 90% Questionnaire Deduced by the students from
Naïve Bayes 77% 76% 76% 86% machine learning manual marking)
Logistic Regression 79% 80% 80% 87% model *100)
K nearest neighbour 76% 78% 78% 90%
Verbal- 39 32 82.05%
Linear discriminant 79% 79% 79% 86%
Linguistic
analysis
Bodily- 27 16 59.26%
Cross-Validation scores
Kinesthetic
Mathematical- 43 36 83.72%
Logical
prediction. Two, the improvement in results in comparison to the pre­ Visual-Spatial 57 52 91.23%
vious researchers. Table 8 and 9 shows the consistency of manual Interpersonal- 34 26 76.47%
marking and machine prediction for Gardner’s theory of Multiple Intrapersonal
Intelligences. Total 200 162 81.00%
Total number of students: 200
Table 8 summarizes the % consistency between manual marking and
machine prediction. Column 2 shows the number of students in each
category of intelligence. As mentioned earlier, we excluded the intelli­ Another view of this research helps to understand the different pat­
gence for which the observation was not possible in an online learning terns of learner behavior during their course of online learning. The
mode. Column 3 has the number of students with dominant intelligence researchers made the following observations during and after the data
by the machine. The last column calculates the % consistency. It can be was collected:
observed that visual-spatial intelligence has the highest consistency. It
also means that the attributes used to identify this intelligence are good. 1. Although learning styles are considered to be static for a learner, and
This is followed by the Mathematical-logical and Verbal-Linguistic close they generally do not change from time to time; During this research,
behind. The bodily-kinesthetic performance has the least consistency; it we have observed some patterns of changes in the learning style of
may mean that the attributes are not enough for prediction; it may also learners. One such remarkable observation was the switch from
mean that the algorithm needs to be fine-tuned more to get better re­ verbal-linguistic to visual-spatial learning style based on the diffi­
sults. This can be taken up as a direction in future work. The overall culty of the learning resource while analyzing learner’s learning
consistency of machine prediction as compared to manual marking is styles using Gardner’s theory of multiple intelligences. Table 10
81%. summarizes the change in the learning style of learners based on the
Table 9 summarizes the % consistency of machine prediction as difficulty of the learning resource.
compared to manual marking for the FSLM learning style model. Values 2. The second observation during this study was the heavy dependence
in the table are compared for each dimension and each style of learning of learners on simulations, case studies, and application-level ques­
in the dimension. For example, the first block consists of the costs for the tions when difficulty level of the concept is high. To observe this
Input dimension. With the manual marking, we had 113 visual students, pattern, we considered a total of 12 concepts, which were explained
and the machine predicted that 101 of them as visual, which is a con­ through 150 learning resources. There were five difficult concepts,
sistency of 89.38%. Eighty-seven students were verbal from their for which the 200 students monitored; 15% of them used other re­
questionnaires, and the machine prediction was 80, the consistency sources, 35% of them preferred case studies, 26% used application-
being 91.95%. The % consistency for all the dimensions ranges between level questions, 24% used simulations. The graph in the figure
80% and 92%, which is acceptable. The consistency was least in the below summarizes this.
Understanding dimension, which may mean that either the model 3. Almost 75% of the learners shifted from intrapersonal to interper­
should be optimized or the attributes should be added during the pre­ sonal learning style, close to assignment submission and assessment
diction of this dimension.

9
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Table 9
% Consistency between manual marking and machine prediction for the FSLM Learning style model.
Dimension Value Number of students with learning style from Number of students with learning style % Consistency ((correct predictions by machine/
Manual Marking of E-Questionnaire Deduced by the machine learning model No of students from manual marking) *100)

Input Visual 113 101 89.38%

Verbal 87 80 91.95%
Total-Input 200 181 90.50%
Processing Active 103 90 87.38%
Reflective 97 85 87.63%
Total-Processing 200 175 87.50%
Understanding Sequential 89 70 78.65%
Global 111 90 81.08%
Total- 200 160 80.00%
Understanding
Perception Sensing 116 103 88.79%
Intuitive 84 75 89.29%
Total-Perception 200 178 89.00%
Average 694 86.75%
consistency
Total number of students: 200

dates. There were a lot of messages being sent on the discussion 2. 221 students out of 498 were initially verbal learners; they preferred
forum on various topics. In the data, during our study, we observed reading material to video lectures and PowerPoint presentations to
that about 75% (374 out of 498) of the learners preferred studying still images or simulations. As the difficulty level of the concepts
themselves or just reading forum messages during the ordinary increase, about 35% of verbal learners shift to visual means of
course of study. But we also observe that as the date of assignment learning by using videos, images, and simulations Fig. 5 shows this.
submission and assessment is due, almost 45% of these students
participate in discussions forums by replying to questions and post­ The methods used in this study are compared to the literature that
ing questions for others to respond. has been cited to bring out the significance of the research we con­
ducted. There are a few parameters on which the judgement was made.
While using the FSLM learning style model for the predictions of Some of the parameters are the consistency of manual marking and
learning styles, the following interesting observations were made: machine prediction, accuracy percentage, patterns in the learning styles,
prediction in all dimensions of Felder Silverman model.
1. There were around 266 students out of 498 students that preferred
the sequential understanding of concepts during the initial predic­ 6. Conclusion and future work
tion of learning styles. They preferred slow and steady progress, not
rushing to complete learning, spending more than 75% of the time In this paper, we have attempted to identify attributes that can be
allotted to study, using the next button more often. But when the used to detect multiple intelligences and also attributes pertaining to the
examination time is near, they use all the resources available for a Felder-Silverman model of learning styles. Some of the attributes are
global learner, using summaries, skimming through reading mate­ time spent in learning through audio, the number of messages sent, and
rial, hastily completing tests. This behavior shows that learning style so on. Data was collected, and algorithms were run to validate the ac­
changes through the course of a learner’s life. curacy of classification. For the data collected to detect the dominating
intelligence of the learner, the highest accuracy was achieved with
Support Vector Machine. The lowest was with a decision tree, and that is
Table 10
45.55%. For the data collected for the FSLM model, the accuracy in each
Summary of Patterns in Change of Learning Style of the Learners Based on the
dimension was calculated for the different algorithms. The highest ac­
Difficulty of Learning Resource.
curacy was achieved when the Support Vector Machine was used in the
Learning Difficulty Type of n(VL n(VS % of learners input dimension. The consistency between manual marking and ma­
resource level resource –> –> with a change
(LR) VS) VL) in learning
chine prediction was also calculated; the results are promising and are
style an improvement in this area. Future work in this area can be feature
LR 1 Easy Fact 1 120 51.05%
LR 2 Easy Definition 2 120 51.48%
LR 3 Intermediate Application of 40 10 21.10%
concept
LR 4 Difficult Case Study 100 2 43.04%
LR 5 Difficult Creating 100 2 43.04%
knowledge
LR 6 Difficult Problem 90 4 39.66%
solving
LR 7 Easy Example 20 80 42.19%
LR 8 Difficult Proof of 80 12 38.82%
theorem
LR 9 Easy Fact 4 90 39.66%
LR 10 Intermediate Definition 40 10 21.10%
LR 11 Intermediate Statement 10 20 12.66%
LR 12 Difficult Apply method 60 16 32.07%
to problem
Number of Learners = 498; n(VL) = 104, n(VS) = 133)
Fig. 5. Summary of Usage of Different Learning Resources when Studying
n(VL): Number of Verbal-Linguistic learners, n(VS): Number of Visual-Spatial Learners
Difficult Concepts.

10
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

reduction and testing the same on a larger dataset. We also observed Carver, C. A., Howard, R. A., & Lane, W. D. (1999). Enhancing student learning through
hypermedia courseware and incorporation of student learning styles. IEEE
various patterns like change in learning styles of the learners according
Transactions on Education, 42(1), 33–38. https://doi.org/10.1109/13.746332
to the change in situations such as exam date nearing; change of learning Cassidy, S., & Eachus, P. (2000). Learning style, academic belief systems, self-report
style based on the difficulty of the concepts. We plan to use the transfer student proficiency and academic achievement in higher education. Educational
learning mode of machine learning to test the validity of the model Psychology, 20(3), 307–322. https://doi.org/10.1080/713663740
Cha, H. J., Kim, Y. S., Park, S. H., Yoon, T. B., Jung, Y. M., & Lee, J.-H. (2006). Learning
proposed in the future by deploying the trained model of the algorithm Styles Diagnosis Based on User Interface Behaviors for the Customization of Learning
to predict the learning style of the learners in a real-world environment. Interfaces in an Intelligent Tutoring System. In M. Ikeda, K. D. Ashley, & T.-
We also plan to generalize the patterns observed in various classes of .-W. Chan (Eds.), Intelligent Tutoring Systems (pp. 513–524). Berlin Heidelberg:
Springer.
learners. A limitation that we will overcome in the future is considering Crockett, K., Latham, A., & Whitton, N. (2017). On predicting learning styles in
non-computer science learners learning online. conversational intelligent tutoring systems using fuzzy decision trees. International
Journal of Human-Computer Studies, 97, 98–115. https://doi.org/10.1016/j.
ijhcs.2016.08.005
7. Availability of data and materials P.Q. Dung A.M. Florea A Literature-based Method to Automatically Detect Learning
Styles in Learning Management Systems Proceedings of the 2Nd International
Conference on Web Intelligence, Mining and Semantics 2012 46:1–46:7. 10.1145/
The datasets generated and analyzed during the current study are not
2254129.2254186.
publicly available. A Non-Disclosure agreement is signed with the uni­ Fatahi, S., Moradi, H., & Farmad, E. (2015). Behavioral feature extraction to determine
versity in which the data was collected. The university is a Central learning styles in E-Learning environments. International Association for Development
University of India in the state of Uttarakhand. However, the data is of the Information Society, 66–72.
Felder, R. M., Silverman, L. K., et al. (1988). Learning and teaching styles in engineering
available from the corresponding author on a reasonable request. The education. Engineering Education, 78(7), 674–681.
trained models are also available on reasonable requests to the authors. Fleming, N. D. (2001). Teaching and learning styles: VARK strategies. IGI global.
García, P., Schiaffino, S., & Amandi, A. (2008). An enhanced Bayesian model to detect
students’ learning styles in Web-based courses. Journal of Computer Assisted Learning,
CRediT authorship contribution statement 24(4), 305–315. https://doi.org/10.1111/j.1365-2729.2007.00262.x
García, P., Amandi, A., Schiaffino, S., & Campo, M. (2007). Evaluating Bayesian
networks’ precision for detecting students’ learning styles. Computers & Education,
Fareeha Rasheed: Conceptualization, Methodology, Data curation, 49(3), 794–808. https://doi.org/10.1016/j.compedu.2005.11.017
Writing - Original draft, Software, Visualization, Investigation. Abdul Gardner, H., & Hatch, T. (1989). Intelligences multiple. Educational Research, 18, Issue 8.
Wahid: Supervision, Writing - review & editing. https://doi.org/10.3102/0013189X018008004
Georgiou, D. A., & Makry, D. (2004). A learner’s style and profile recognition via fuzzy
cognitive map. IEEE International Conference on Advanced Learning Technologies,
2004. Proceedings., 36–40. 10.1109/ICALT.2004.1357370.
Declaration of Competing Interest Graf, S., Liu, T.-C., & Kinshuk.. (2010). Analysis of learners’ navigational behaviour and
their learning styles in an online course. Journal of Computer Assisted Learning, 26(2),
116–131. https://doi.org/10.1111/j.1365-2729.2009.00336.x
The authors declare that they have no known competing financial
Graf, S. (2006). An Approach for Detecting Learning Styles in Learning Management
interests or personal relationships that could have appeared to influence Systems. Sixth IEEE International Conference on Advanced Learning Technologies,
the work reported in this paper. ICALT’06, 2–4.
Hasibuan, M. S., Nugroho, L. E., & Santosa, P. I. (2019). Model detecting learning styles
with artificial neural network. Journal of Technology and Science Education, 9(1),
Acknowledgement 85–95.
Hmedna, B., Mezouary, A. El, & Baz, O. (2017). Identifying and tracking learning styles
in MOOCs: A neural networks approach. Advances in Intelligent Systems and
Not applicable. Computing, 520(2), 125–134. https://doi.org/10.1007/978-3-319-46568-5_13
Hmedna, B., Mezouary, A. El, Baz, O., & Mammass, D. (2016). A Machine Learning
Approach to Identify and Track Learning Styles in MOOCs. In 2016 5th International
Appendix A. Supplementary data
Conference on Multimedia Computing and Systems (ICMCS) (pp. 212–216).
Honey, P., & Mumford, A. (1989). Learning styles questionnaire. Incorporated:
Supplementary data to this article can be found online at https://doi. Organization Design and Development.
org/10.1016/j.eswa.2021.114774. C.-W. Hsu C.-C. Chang C.-J. Lin et al. A practical guide to support vector classification
2003 Taipei.
Huang, E. Y., Lin, S. W., & Huang, T. K. (2012). What type of learning style leads to
References online participation in the mixed-mode e-learning environment? A study of software
usage instruction. Computers & Education, 58(1), 338–349. https://doi.org/10.1016/
j.compedu.2011.08.003
Aissaoui, O. E., El, Y., El, A., Oughdir, L., & Allioui, Y. E. (2019). A fuzzy classification
Igbrue, C., & Pathak, P. (2008). A framework for Creating Multiple Intelligences
approach for learning style prediction based on web mining technique in e-learning
Informed Content for e-learning. In C. J. Bonk, M. M. Lee, & T. Reynolds (Eds.),
environments. Education and Information Technologies, 24, 1943–1959.
Proceedings of E-Learn: World Conference on E-Learning in Corporate, Government,
Alkhuraiji, S., Cheetham, B., & Bamasak, O. (2011). Dynamic Adaptive Mechanism in
Healthcare, and Higher Education 2008 (pp. 1657–1666). Association for the
Learning Management System Based on Learning Styles. In 2011 IEEE 11th
Advancement of Computing in Education (AACE). https://www.learntechlib.org/p/
International Conference on Advanced Learning Technologies (pp. 215–217). https://
29876.
doi.org/10.1109/ICALT.2011.69
Ignacio, N., Rodr, L., Lugo, G. S., Castro, L. A., & Kono, M. D. (2015). A Framework for
Angeli, C., Valanides, N., Polemitou, E., & Fraggoulidou, E. (2016). An interaction effect
Automatic Identification of Learning Styles in Learning Management Systems., 106,
between young children’s field dependence-independence and order of learning with
59–68.
glass-box and black-box simulations: Evidence for the malleability of cognitive style
Jegatha Deborah, L., Baskaran, R., & Kannan, A. (2014). Learning styles assessment and
in computer-supported learning. Computers in Human Behavior, 61, 569–583. https://
theoretical origin in an E-learning scenario: A survey. Artificial Intelligence Review, 42
doi.org/10.1016/j.chb.2016.03.060
(4), 801–819. https://doi.org/10.1007/s10462-012-9344-0
Azzi, I., & Jeghal, A. (2019). A robust classification to predict learning styles in adaptive
Kaewkiriya, T., Utakrit, N., & Tiantong, M. (2016). The design of a rule base for an e-
E-learning systems. Education and Information Technologies, 25, 437–448.
Learning recommendation system base on multiple intelligences. International
Bernadó-Mansilla, E., & Garrell-Guiu, J. M. (2003). Accuracy-based learning classifier
Journal of Information and Education Technology, 6(3), 206–210. https://doi.org/
systems: Models, analysis and applications to classification tasks. Evolutionary
10.7763/IJIET.2016.V6.685
Computation, 11(3), 209–238.
Kearns, M., & Mansour, Y. (1999). On the boosting ability of top–down decision tree
Bernard, J., Chang, T. W., Popescu, E., & Graf, S. (2017). Learning style Identifier:
learning algorithms. Journal of Computer and System Sciences, 58(1), 109–128.
Improving the precision of learning style identification through computational
Khan, M. S. H., Shamim, M. R. H., & Nambobi, M. (2018). Learning styles and online
intelligence algorithms. Expert Systems with Applications, 75, 94–108. https://doi.
tools: How to construct an effective online learning environment. In Optimizing
org/10.1016/j.eswa.2017.01.021
student engagement in online learning environments (pp. 147–162). IGI Global.
Bishop, C. M. (2006). Pattern recognition and machine learning. Springer.
Kolås, L., & Staupe, A. (2007). A personalized e-learning interface. In EUROCON 2007 -
Bursac, M., Milosevic, D., & Mitrović, K. (2019). Proposed Model For Automatic Learning
The International Conference on Computer as a Tool (pp. 2670–2675). https://doi.org/
Style Detecting Based On Artificial Intelligence. 11th International Scientific
10.1109/EURCON.2007.4400362
Conference “Science and Higher Education in Function of Sustainable Development“.
Kolb, A. Y., & Kolb, D. A. (2005). Learning styles and learning spaces: Enhancing
Carmona, C., Castillo, G., & Millán, E. (2008). Designing a dynamic bayesian network for
experiential learning in higher education. Academy of Management Learning &
modeling students ’ learning styles. Eighth IEEE International Conference on Advanced
Education, 4(2), 193–212.
Learning Technologies, 2008, 346–350. https://doi.org/10.1109/ICALT.2008.116

11
F. Rasheed and A. Wahid Expert Systems With Applications 174 (2021) 114774

Kolekar, S. V., Sanjeevi, S. G., & Bormane, D. S. (2010). Learning style recognition using Rasheed, F., & Wahid, A. (2019). In Learning Style Recognition: A Neural Network Approach
Artificial Neural Network for adaptive user interface in e-learning. IEEE International (pp. 301–312). Springer Singapore.
Conference on Computational Intelligence and Computing Research, 2010, 1–5. https:// Rish, I. (2001). An empirical study of the naive Bayes classifier. IJCAI 2001 Workshop on
doi.org/10.1109/ICCIC.2010.5705768 Empirical Methods. Artificial Intelligence, 3(22), 41–46.
Komarraju, M., Karau, S. J., Schmeck, R. R., & Avdic, A. (2011). The Big Five personality Sangineto, E., Capuano, N., Gaeta, M., & Micarelli, A. (2008). Adaptive course generation
traits, learning styles, and academic achievement. Personality and Individual through learning styles representation. Universal Access in the Information Society, 7
Differences, 51(4), 472–477. https://doi.org/10.1016/j.paid.2011.04.019 (1), 1–23. https://doi.org/10.1007/s10209-007-0101-0
Kotsiantis, S. B., Zaharakis, I., & Pintelas, P. (2007). Supervised machine learning: A Santo, S. A. (2006). Relationships between learning styles and online learning.
review of classification techniques. Emerging Artificial Intelligence Applications in Performance Improvement Quarterly, 19(3), 73–88. https://doi.org/10.1111/j.1937-
Computer Engineering, 160(1), 3–24. 8327.2006.tb00378.x
Latham, A., Crockett, K., McLean, D., & Edmonds, B. (2012). A conversational intelligent Sheeba, T., & Krishnan, R. (2019). Automatic detection of students learning style in
tutoring system to automatically predict learning styles. Computers & Education, 59 learning management system. Springer International Publishing. https://doi.org/
(1), 95–109. https://doi.org/10.1016/j.compedu.2011.11.001 10.1007/978-3-030-01659-3
Lynch, T. G., Woelfl, N. N., Steele, D. J., & Hanssen, C. S. (1998). Learning style Sternberg, R. J. (1994). Allowing for thinking styles. Educational Leadership, 52(3),
influences student examination performance. The American Journal of Surgery, 176 36–40.
(1), 62–66. https://doi.org/10.1016/S0002-9610(98)00107-X Sutton, O. (2012). Introduction to k nearest neighbour classification and condensed nearest
Mankad, K., Sajja, P. S., & Akerkar, R. (2011). An automatic evolution of rules to identify neighbour data reduction (pp. 1–10). University of Leicester: University Lectures.
students multiple intelligence. Communications in Computer and Information. Science, Vemuri, V. R. (1993). Main problems and issues in neural networks application. In
133 CCIS(PART 3), 35–45. https://doi.org/10.1007/978-3-642-17881-8_4 ANNES 1993–1st New Zealand International Two-Stream Conference on Artificial Neural
Marsland, S. (2015). Machine learning: An algorithmic perspective. CRC Press. Networks and Expert Systems (p. 226). https://doi.org/10.1109/ANNES.1993.323037
Nakayama, M., & Santiago, R. (2012). Learner characteristics and online learning. Vermunt, J. D., & Vermetten, Y. J. (2004). Patterns in student learning: Relationships
Encyclopedia of the Sciences of Learning, 1745–1747. between learning strategies, conceptions of learning, and learning orientations.
Özpolat, E., & Akar, G. B. (2009). Computers & Education Automatic detection of Educational Psychology Review, 16(4), 359–384.
learning styles for an e-learning system. Computers & Education, 53(2), 355–367. Villaverde, J. E., Godoy, D., & Amandi, A. (2006). Learning styles’ recognition in e-
https://doi.org/10.1016/j.compedu.2009.02.018 learning environments with feed-forward neural networks. Journal of Computer
Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and Assisted Learning, 22(3), 197–206. https://doi.org/10.1111/j.1365-
evidence. Psychological Science in the Public Interest, 9(3), 105–119. 2729.2006.00169.x
Pitigala Liyanage, P., Gunawardena, L., & Hirakawa, M. (2016). Detecting Learning Vincent, A., & Ross, D. (2001). Personalize training: Determine learning styles,
Styles in Learning Management Systems Using Data Mining. Journal of Information personality types and multiple intelligences online. The Learning Organization, 8(1),
Processing, 24, 740–749. https://doi.org/10.2197/ipsjjip.24.740 36–43.
Popescu, E. (2009). Diagnosing Students’ Learning Style in an Educational Hypermedia Wild, C. J., Pfannkuch, M., Tomlinson, C. A., Simister, C., Protheroe, N., Nickerson, R. S.,
System. In Cognitive and Emotional Processes in Web-Based Education: Integrating … Persson, T. (2014). Web-Based Education. Review of Educational Research, 85, Issue
Human Factors and Personalization. 10.4018/978-1-60566-392-0.ch011. 2. https://doi.org/10.1080/13670050608668629
Racelis, R., & Iii, M. (2016). Classification of Learning Styles in Virtual Learning Zhang, K., & Bonk, C. (2009). Addressing diverse learner preferences and intelligences
Environment using Data Mining : A Basis for Adaptive Course Design. 56–61. with emerging technologies: Matching models to online opportunities. Canadian
Rajper, S., Shaikh, N. A., Shaikh, Z. A., & Mallah, G. A. (2016). Automatic detection of Journal of Learning and Technology / La Revue Canadienne de l’apprentissage et de La
learning styles on learning management systems using data mining technique. Indian Technologie, 34. https://doi.org/10.21432/T2530K
Journal of Science and Technology, 9(15). https://doi.org/10.17485/ijst/2016/v9i15/ Zhang, L. F. (2011). The malleability of intellectual styles. In The Malleability of
85959 Intellectual Styles (Vol. 9781107096448)Cambridge University Press. https://doi.
org/10.1017/CBO9780511973055.

12