See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/379478419

Predict student learning styles and suitable assessment methods using click
stream

Article in Egyptian Informatics Journal · June 2024

DOI: 10.1016/j.eij.2024.100469

CITATIONS READS

2 116

4 authors, including:

Mohamed Khafagy Marwa Hussien Mohamed

Fayoum University October 6 University
85 PUBLICATIONS 767 CITATIONS 11 PUBLICATIONS 189 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Marwa Hussien Mohamed on 02 April 2024.

The user has requested enhancement of the downloaded file.

 Egyptian Informatics Journal 26 (2024) 100469

Contents lists available at ScienceDirect

Egyptian Informatics Journal

journal homepage: www.sciencedirect.com

Predict student learning styles and suitable assessment methods using

click stream
Ahmed Rashad Sayed a, b, Mohamed Helmy Khafagy c, Mostafa Ali b, Marwa Hussien
Mohamed a, *
a
Information System Department, Information System and Computer Science Faculty, October 6 University, 6th of October 12573, Egypt
b
Information System Department, Faculty of Computers and Artificial Intelligence, Fayoum University, Faiyum 63514, Egypt
c
Computer Science Department, Faculty of Computers and Artificial Intelligence, Fayoum University, Faiyum 63514, Egypt

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

Keywords: Adaptive learning, which aims to give each learner engaging, effective learning experiences, is one method of
Adaptive learning offering modified education. Adaptive learning seeks to consider the student’s unique characteristics by
Machine learning personalizing the learning course materials and evaluation procedures. To determine the student’s preferred
Classification
learning strategies, we first ascertain their attributes utilizing VAK learning styles. In this study, we developed an
Learning styles
integrated model to classify learners based on their learning activity clicks by combining machine learning al­
gorithms like K-Nearest Neighbor (KNN), random forest (RF), and support vector machine (SVM) and Logistic
regression (LR) with semantic association, which is used to help us map learning activity with VAK learning style.
This enables us to classify learners, determine their preferred methods of learning, and offer the most suitable; as
a result, we were able to group pupils according to their learning styles and provide the best evaluation technique
or strategies. To assess the effectiveness of the suggested model, several tests were executed on the actual dataset
(Open University Learning Analytics Dataset, or OULAD). According to studies, using a Random Forest algo­
rithm, the suggested model can predict which evaluation strategy or strategies will be most effective for each
student and can classify individuals with the highest degree of accuracy—98%.

1. Introduction different levels of knowledge among students during the study process.
The concept of learning style is the way students choose to study effi­
The development of the educational system because of the ciently. Students will understand their wishes during the learning pro­
advancement of Information Technology, particularly the Internet, cess if they are familiar with their favourite learning method. Since
creates prospects for improving the information services provided by Moodle is unable to detect students’ learning preferences automatically.
educational institutions. E-learning is an educational method that in­ Student behaviour should be analyzed based on the number of times
volves technological tools. One of the e-learning skills is the Learning they access e-learning in Moodle, and completing a questionnaire on
Management System (LMS). Students can enter lecture data, discussion learning styles is required to determine the best learning style for each
boards and chat rooms and access lecture assignments given by lecturers student [2].
[1]. Through the Learning Management System. The popular learning Students can access video lectures and educational resources on the
management system that is now in use is Moodle. Students’ skills and leading MOOC platforms. In addition to successful completion of the
actions during online learning can be recorded by Moodle and kept in course, students will also be awarded a certificate. The MOOC platform
the logs of Moodle. ensures that the material is free from grammar, punctuation, and
The problematic with e-learning is that undergraduates are less spelling mistakes. Although MOOCs are very popular, the high dropout
engaged and leave the classroom more frequently; therefore, it is rate of students and poor performance measures are frustrating, as less
necessary to understand students’ preferences in the learning process by than 13 % of pupils [3]. As a result, significant contributors to students’
considering the learning style of each situation student. where there are lack of interest in courses include their lack of participation and their

* Corresponding author.
E-mail addresses: [email protected], [email protected] (A. Rashad Sayed), [email protected] (M. Helmy Khafagy), mam16@fayoum.
edu.eg (M. Ali), [email protected], [email protected] (M. Hussien Mohamed).

https://doi.org/10.1016/j.eij.2024.100469
Received 28 December 2023; Received in revised form 14 March 2024; Accepted 28 March 2024
1110-8665/© 2023 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Computers and Artificial Intelligence, Cairo University. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

difficulties in tracking resources and activities for assessments [4]. developed [8]. Learner personality, which are the traits of the learner
Therefore, in online education systems, student participation is a that influence the learning process [9], is a more general term that in­
crucial element in the success of the course. Although a virtual learning cludes learning varieties. In the context of learning, the learning style is
environment with a certification of achievement is known as a MOOC, a fascinating characteristic. The chosen method of employing one’s ca­
the online learning platform is assumed to be more of a hybrid [5]. pacity for learning is their learning style. Different psychologists have
Student satisfaction and the calibre of the educational experience are presented us with several learning style models.
mainly related to student participation. To reduce the dropout rate, it is There were four types of learners: activists, theorists, pragmatists,
of great importance to realize how students interact with the interests of and reflectors. The learning modalities describe the various ways that
continuing professional education. students can learn. The four modalities that are typically considered are
The improved learning environment helps to promote self-control tactile (touching), kinesthetic (moving), auditory (hearing), and visual
and motivation, which helps students stay on task and perform better. (seeing). Visual aid learners frequently learn by viewing things, whereas
Understanding students, learning, and interacting with virtual learning auditory learners love reading or listening. The tangible learns best
environments are essential to creating better learning environments [6]. when they touch items and experience their configuration and sensa­
The perfect learning models and the customization of learning envi­ tion, while the kinesthetic learns best by doing rather than merely
ronments depend greatly on student participation. watching [10].
Research groups use learning analytics to predict student perfor­ Linda Silverman and Richard Felder created the Felder-Silverman
mance due to the evolution of educational data. The current study uses Index of Learning Styles (ILS) in 1988. This model analyses learning
datasets from the Open University Learning Analytics dataset collected styles and makes four-dimensional preference distinctions. visual/ver­
by Flea at the Open University in the UK to predict student participation. bal, sensing/intuitive, active/reflective, and sequential/global. These
Compared to MOOC platforms, virtual learning environments (VLEs) four dimensions correspond to four psychological processes: input,
enable instructors to monitor student actions and assessments because perception, processing, and understanding.
they were designed to move passive learners toward active learning. As a We can implement the VAK (Visual, Auditory, Kinesthetic) mapping
result of the VLE’s recording of student interaction, the instructor can process with teachers in classrooms by detecting the students’ learning
better understand how students behave in these learning environments. styles by using various methods like surveys, questionnaires, and ob­
The data of these learning platforms have made it possible to make servations. We will apply this step first of the course teaching this part 1
decisions that are consistent with the data, and the clickstream data of then we will prepare the material and discuss it with the students in the
these platforms can even predict student engagement at an early stage. classroom based on the learner style we determine.
The data set includes information on demographics and student in­ For visual students, we will add posters, charts, graphs, and multi­
teractions with the VLE. It is laborious for anyone to use conventional media presentations to demonstrate concepts, processes, and relation­
approaches to examine and produce valuable information because the ships. For auditory students, we will add lectures, conversations, audio
data is kept in the Open University’s extensive data repository [7]. recordings, and oral presentations. Also, encourage students to engage
Research goal: in group discussions, debates, and presentations to improve their audi­
tory learning experiences.
• This work used classification approaches, such as machine learning For kinesthetic students, we will provide hands-on learning experi­
(ML) techniques, to construct a model to analyze student interaction, ences, such as experiments, demonstrations, simulations, and role-
predict learning styles, and appropriate evaluation methods in VLE playing. Table 1 shows the relation between a list of tools and their
courses. learning style orientations (Visual, Auditory, Kinesthetic).
• Learning analytics frequently employs classification approaches. The Teachers can develop inclusive learning environments that accom­
aim is to investigate how VLE activities affect students’ participation. modate students’ different learning styles and preferences by incorpo­
• The novel of this work is to create a machine learning model that can rating the VAK mapping process into their classroom instruction. This
detect a learner’s involvement in their studies at an early point of the strategy encourages student involvement, motivation, and academic
course based on the number of clicks activity in the course. performance by personalizing instruction to individual needs and
• The main idea is to increase student involvement to help instructors interests.
develop appropriate strategies. In this work, we explore an ideal
collection of variables that reliably predict the engagement of the 3. Machine learning
learners.
• Based on the students’ clicks on their learning activities, our sug­ Prediction and classification involve machine learning models and
gested model can classify students with an average accuracy of 98 % algorithms. Machine learning is the convergence of statistics and com­
and predict the best evaluation techniques for each student. puter science, in which machines learn to improve performance from
previous experience, just as humans. The only difference is that
In this paper, we have organized our discussion into several sections.
Section 2 elaborates on the learning style, while Section 3 lists machine Table 1
learning techniques. In Section Four, we present some related work, Relation between a list of tools and their learning style orientations (Visual,
followed by a discussion of student modelling in Section Five. Section Auditory, Kinesthetic).
Six outlines our research question, while Section Seven proposes a Tool Learning Style Category
model and methodology. In Section Eight, the experimental results, and
posters Visual
in Section Nine, we recommend an assessment method. Finally, we
charts
conclude with a summary of our findings and outline areas for future Graphs
work. multimedia presentations
lectures Auditory
oral presentations
2. Learning style
audio recordings
conversations
Since a few decades ago, personalization has been a feature of experiments Kinesthetic
computing, and all current systems provide users with a personalized demonstrations
experience. From direct-programmed training and examinations to simulations
role-playing
adaptable virtual environments, e-learning systems have significantly

2
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

computers learn from data, while humans learn through events and engage with the VLE, their scores, and the outcomes. A reduced feature
experiences. Machine learning includes three types of learning algo­ vector is created using these variables as the foundation. The linear
rithms: supervised, unsupervised, and reinforcement. regression model served as the baseline of the study. The Random Forest
Supervised learning occurs when data, along with labels, are sent to a classification technique produced the best findings of the model, which
computer as input. The machine learns patterns from it and tries to were 95 % accuracy, 95 % exact, and 98 % appropriate. The click ac­
predict the label of the newly provided data. For example, the computer tivities that serve as an efficient interface between students and the VLE
will estimate whether it will rain today or not based on temperature, are a subset of click activities that are pertinent to early prediction.
humidity, precipitation, and other weather factors, after learning the In [16], this study creates a robust prediction model is created and
data [11]. essential variables are identified that significantly impact student effects
In unsupervised learning without labels, the computer must examine in an e-learning environment. The significant contribution consists of
patterns in the data and classify them based on the properties of the data. two parts: highlighting some experimental visions under the impact of a
For example, tweets are automatically tagged according to different collection of factors using feature selection techniques and proposing a
topic categories [12]. prediction model using the most important characteristics while using
Classification in machine learning is used to create forecasts after the the K-fold cross-validation approach. Using student data from the
machine has been educated from the data. There are many classification Learning Management System, it is thoroughly examined how various
applications, from serious applications, such as disease prediction and variables affect model performance and how input and target output
traffic monitoring, to simple applications, such as identifying spam and correlations relate to each other. The most common machine learning
games. A naive Bayes classifier is a set of classification methods based on techniques are then evaluated against the suggested approach. The
Bayes theorem. The algorithms share common elements such that each findings showed that while less engaged students in the course tend to
pair of classified attributes is self-dependent on each other [13]. interact less frequently, they perform much better when using the e-
Support vector machine (SVM) is based on the idea of defining a learning system [17]. Additionally, according to the study findings,
hyperplane that best divides a data set when displayed on a graph into some prediction systems, like the Random Forest method, have signifi­
two classes. The K-Nearest Neighbors (KNN) algorithm uses “similarity” cant advantages in predicting student performance and can do so with
to predict the value of new data points, meaning that the new data point an accuracy rate of up to 80 %. There is also a discussion of other student
will be assigned a value based on its similarity. features that could be useful in the e-learning system.
Using AI in education to determine VAK learning styles can provide In [18], this study collects VLE data and processes using many pre­
various advantages, such as personalized learning experiences, tailored processing algorithms, such as removing missing values, normalization,
instructional tactics, and increased student engagement. Here’s how AI encoding, and outlier identification. Several machine learning (ML)
can recognize VAK learning patterns in education: AI algorithms may classification algorithms were applied to our data, and each algorithm’s
analyze a range of data, including student performance, interactions performance was evaluated using cross-validation techniques and a
with learning materials, and assessment results. variety of useful indicators. Metrics, including precision accuracy,
AI can predict visual, auditory, and kinesthetic learning styles by recall, and AUC (Area Under Curve) scores, are used to assess the
analyzing patterns in student behaviour and learning outcomes. AI- model’s performance. The findings indicate that the CAT (Computerized
powered interactive learning platforms can tailor information and ac­ Adaptive Testing) Boost model is more accurate than the others. When
tivities to individual learning methods. comparing this new model to earlier research, it performed better in all
Using AI technology to detect VAK learning styles in education al­ respects.
lows instructors to develop more personalized and engaging learning According to the Results section of this work, the CAT Boost model
experiences that match the different needs of their students. has an AUC score of 96.24 %, an accuracy of roughly 92.23 %, a pre­
cision of 94.40 % and a recall rate of 100 %. Overall performance was
4. Related work similar for the multilayer perceptron, random forest, and Boost predic­
tive models. We tested our model to the AISAR (Artificial Intelligence-
In [14], to verify the effect of engagement on pupil performance, Based Student Assessment and Recommendation System) model and
machine learning (ML) algorithms were used in this study to detect found that our results were more accurate, with a 94.64 % accuracy
little-engagement learners in social science classes at the Open Univer­ compared to the AISAR model’s 91 % accuracy. Compared to our
sity (OU)., The highest educational level, outcomes, assessment score models, which had a recall of about 92 %, the AISAR model had only
and the number of clicks on virtual learning environment (VLE) activ­ about 50 %.
ities, such as a glossary, homepage, forum, collaborate, content, re­
sources, subpages, data plus and URL during the first-course assessment, 5. Student modeling
were among the study’s input variables. The degree of student partici­
pation in the various activities was the output variable [15]. Some ML It can be challenging to predict how students will behave in learning
algorithms were applied to the data set to predict low-engagement situations. Starting with raw OULAD data, the data preparation step
students. These techniques were used to create training models which produced a refined and condensed data set that was used to create a
were then compared for accuracy and kappa values. The results showed prediction model. The classification techniques used in this work are
that the J48 algorithm consumed an accuracy of 88.5 %, the decision KNN, Random Forest, and Decision Tree. The adoption of these classi­
tree gained an accuracy of 85.9 %, JRIP had an accuracy of 83.27 %, and fication techniques led to their selection.
gradient-boosted classifiers had an accuracy of 86.43 %. Researchers in educational data mining have successfully applied
In [7] this study, data collected over multiple years are used to them and have obtained an extreme performance model utilizing many
predict student participation in the initial stages of a virtual learning methods [19]. The decision tree is a classification approach that can be
environment (VLE). A portion of the Open University Learning Analytics used to extract student behavior patterns from educational data. The
Dataset, published by the British Open University (OU), is given to make Decision Tree methodology needs less data cleansing and is unaffected
predictive models using a machine learning approach. The study’s data by missing values and outliers. The consequence of the logistic regres­
set includes 7,775 undergraduates who assumed social science classes sion method decides the predicted likelihood of a collaborative exclusive
over numerous evaluation years. Experiments are carried out with a set event coming (to) based on multiple external inputs.
of restricted characteristics to determine whether students are actively An ensemble classifier, such as Random Forest, combines the find­
participating in their courses at a high or low level. The most critical ings of various decision trees. According to KNN, related things are al­
factors in predictive analysis are the features that show how students ways close together, but Random Forest is slower than decision trees and

3
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

requires a lot more computing power. This explains why different al­ styles.
gorithms were picked to develop the predictive model. These algorithms The shift from a four-modality approach (tactile, kinesthetic, audi­
have been developed by researchers using Python libraries. Pandas, tory, and visual) to a three-modality approach (kinesthetic, auditory,
NumPy, and Sklearn are the packages used. The student training dataset, and visual) can be explained by shared traits and conceptual clarity.
which makes up about 80 % of the dataset, is available for the classifiers Here are some of the causes behind this transformation: Tactile and
to learn from. kinesthetic modalities frequently overlap in practice. Tactile learning
uses the sense of touch and is closely related to kinesthetic learning,
6. Research question which involves physical motion and interaction. Many tactile-based
activities include kinesthetic movement. As a result, merging both mo­
In this research, to predict student learning styles in a virtual dalities into a single (kinesthetic) simplifies classification and eliminates
learning environment, we compare various machine-learning ap­ redundancy.
proaches. Instructors can use the results for the prediction and accuracy
of our model at an early stage in studying students for the course. Every 7.1. Using the VAK learning style to map student learning activities
student has many characteristics that make it difficult for the models to
make compelling predictions. This research was used to find the best In [20], this stage entails applying a map between student data and
model to classify the learning style of the students using data from the the VAK form. The mapping process makes use of semantic technology.
VLE log. We will answer two questions: In the suggested model, the calculation of the semantic relationship with
RQ1: which is the best model for good performance to predict the student input and VAK classifications makes use of Wordnet, a reputable
learner style? source for linguistic comprehension, as well as the benefit of semantic
RQ2: Which is the best examination (assessment) model for students similarity methods. As a result, the mapping procedure is broken down
based on their learning style? into two parts listed below. Building a corpus of VAK study methods for
learners is the first stage.
7. Methodology The student’s current activities are then mapped out. In this work, an
example study of data from Open University is provided. However, it is
The suggested methodology for advising students on a suitable essential to note that the mapping technique is typically applicable to all
learning style and evaluation strategy based on their learning prefer­ student activities.
ences is described in this section. The architecture comprises three
essential elements, as shown in Fig. 1. The first step is to use various 7.1.1. Constructing corpus of VAK studying model
mapping techniques to enter student data in the VAK form. Based on In [21] claims that beginning in the 1920s, psychologists and
their data, students will be classified as interested in any of the three educational experts created the first VAK principles. The VAK model is
VAK learning modes in the second half. Fig. 1 explains how the repre­ now offered by elaborating on a user’s preference regarding a particular
sentation splits into three main parts. text description. As a result, the tokens that stand for nouns are extracted
Various mapping techniques primarily map Students’ data into the from the VAK text and any unnecessary words are removed.
VAK form. Based on their data, the next component aims to categorize This procedure creates three vectors, each representing one of the
the students into one of the three (visual, auditory, or kinesthetic) VAK three learning styles and listing the keywords that best describe it. Then,
study methods. The third and final component forecasts the most a learning-style dictionary was created. Following the tokenization
effective assessment methods based on the learners’ preferred learning procedure, WordNet was used to create a dictionary with synonyms and

Fig. 1. Proposed Model.

4
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

related words for the newly created learning style keywords. In this step, VAK category, it will be put into that category in this case.
WordNet capabilities to find synonyms and associated phrases are uti­ In other words, the result of the semantic similarity assessment will
lized. Constructed on the semantic similarity relationship involving represent how closely a given learning activity reflects a given learning
models in an identical context, the get related words function gets style. The learning activity will be mapped to the learning style group
associated terms. At the end of this step, the vector for the individual using the semantic similarity metric and a predetermined threshold.
education style grouping would then be a template that includes all the We use a few different equations to figure out how closely two
alternative expressions and related words that best describe that type of different terms are related to semantic similarity. We use I.C., path and
study category. depth, and a hybrid measure to determine their proximity. We also
This stage makes use of WordNet’s ability to locate synonyms and factor in the LCA of both senses and how deep they are in the taxonomy.
related terms. The obtain-associated-words function returns phrases
Similarity (C1, C2) = (2 × N3) / (N1 + N2 + (2 × N 3)) (1)
associated with the models and is built on the semantic similarity rela­
tionship between models in the same context. After this process, a matrix Where:
that includes all alternative expressions and related words that best
describe that category of studying type would then serve as the vector • N1, the minimal collective super notion of C1 and C2, is the length
for the individual education stylishness grouping. specified as several nodes on the path from C1 to C3.
• N2 represents the number of nodes along the path from C2 to C3 in
7.1.2. Mapping process terms of length.
With this strategy, educational activities for learners are mapped to • N3 represents the hierarchy’s overall depth and acts as a scaling
VAK learning-type groups. There are two fundamental categories into factor.
which the mapping process can be split: full matching and semi-
matching. This stage creates the mapping between educational actions Wu and Palmer Similarity (Wu and Palmer) Measure (WPM) is used
in the dataset and the VAK studying type dictionary using semi- and full in the proposed Semantic Sentence (SST) similarity calculation meth­
matching. From the individual study activities in the collection, the odology [23]. The WPM quantifies the similarity between two words
noun-representing tokens are first taken out. All learning activity results based on a score. It is based on the complexity of the two senses within
are kept as a vector that can be compared to the dictionary of education the taxonomy, as well as the complexity of their (LCS). (Farthest spec­
type classifications and the mapping. ified ancestor node). Note that the shortest path connecting these two
senses is not usually the least common subsumer (LCS), which is the
7.1.2.1. Full matching category. In [22], the index terms in this category deepest common ancestor within the taxonomy. The node closest to
were sorted into a list of synonyms using careful matching based on these two perceptions is not it. When there are multiple candidates for
WordNet synonym relations. In this mapping type, synonyms for the LCS (the LCS has more than one path to the root), the longest path is
learning style traits also include descriptions of learning activities. This used to determine similarity. Wu and Palmer calculated the degree of
establishes a strong correlation connecting the interest and the subse­ semantic similarity between concepts C1 and C2 using Eq. (1).
quent studying from 100 % of mapping, producing an exact match when The suggested approach uses the Wu- and Palmer-based WordNet
using a dictionary-based mapping. Each token from the education functions to assess how much two concepts within the same information
activity’s description vector is contrasted with the Python learning style content (I.C.) resemble each other. See the following section for further
dictionary throughout the comparison step. information on the Wu and Palmer-based WordNet function, which
There are two circumstances. Most of the learning activity tokens in determines the degree of similarity between two words. The similarity
the first are grouped under a single learning style category. This activity score in our system indicates how closely connected the activity
is suited to the pattern categorization study in this situation. As of the description and the learning style keywords are to one another.
second in this case, this activity is linked explicitly to the pattern cate­ The similarity score was obtained using the formula:
gorization study. If the educational activity tokens are present in two or Wordnet.wup_similratiyScore(“synset(’keywords.pos.nn’)”, “words.
more categories with an equal quantity of tokens and the activity is pos.nn”).
mapped to both categories in the second scenario. The OULAD datasets have many activities Table 2 lists some activ­
ities on the VLE logs datasets and its description.
7.1.2.2. Semi-matching category. In [22], Semantic similarity served as Following of the mapping technique, each learning activity that the
the basis for creating this category. Semantic similarity is a way of learner starts is assigned to one or more of the VAK learning styles based
describing the link between two models that have similar attributes. on the activity belonging to full matching or semi-matching. This
Using a specific threshold, this category would then include ideas mapping identifies a particular VAK learning style group in which this
loosely related to the original idea. The topological parameters edge and educational activity fits.
information content (I.C.) inside the taxonomy enable us to measure or We found that the three activities shared_subpages, subpage, and
evaluate the semantic similarity of concepts in accordance with Tver­ ou_elluminate have a semi-matching in our coding we selected the
sky’s cognitive psychology hypothesis at this level. highest percentage as shown in the results in Table 3. As a result, it
A few examples are the hierarchy created by “is-a” relations, the
member/member-of relations with the taxonomic importance, and fea­
Table 2
tures based on the has-part connection. Therefore, if there is a “is-a,”
Activity description.
“has-part,” or “member of” relationship between the activity and a
learning style category at this level in the WordNet ontology, then the Learning Activity Description
Activity
activity is related to that learning style category.
Forming Use of discussion forums, where learners and teachers can engage
in conversations and facilitate the exchange of ideas.
7.1.2.3. Semantic similarity measure calculations. In [22], if none of the
Dualpane Split the display into two sections (left and right) to allow learners
learning activities perfectly fits into any of the categories of learning to choose the type of exercise or material they want to add.
styles. The semantic similarity measure would be applied during the Ouwiki The educational wiki is a web page in HTML where learners and
similarity mapping process to establish how closely two concepts within teachers can contribute remarks or alter material to improve
the same information content (I.C.) are related to one another. interaction and cooperation in the learning environment.
Ou collaborates video conferencing
Depending on how much the learning activity resembles a particular Ou elluminate Audio-only conferencing

5
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

Table 3 presentations or diagrams, whereas auditory learners may spend

Result of mapping between VAK learning style categories and learning activities. more time listening to audio lectures or debates.
Activity Name Visual Score Auditory Score Kienthatic Score
The OULAD dataset offers a variety of educational data, such as user
glossary 0 0 1
dualPane 0 0 1 click behaviours, style information, and other essential characteristics.
dataplus 0 0 1 Analysing the relationship between user click actions and style data can
glossary 0 0 1 provide useful insights into students’ learning preferences and behav­
shared_subpages 0.66 0.63 0.83 iours, influencing instructional design, personalized learning activities,
subpage 0.66 0.53 0.71
questionnaire 1 0 0
and student support techniques in online education.
ou_collaborate 1 0 0 The original datasets didn’t contain a learning style. We make a
ou_content 1 0 0 preprocessing step to enhance the datasets based on the student’s ac­
repeat_activity 1 0 0 tivity on VLE to map between the student’s activity and VAK learning
resource 1 0 0
style this shown in the previous section.
URL 1 0 0
HTML activity 0 1 0 As an example, ou_content, questionnaire and repeat_activity belong
ou_wiki 0 1 0 to visual and htmlactivity and forming belongs to Auditory.
forming 0 1 0
htmlactivity 0 1 0
8.2. Preprocessing of data
ou_elluminate 0.72 0.8 0.6

Seven relational tables make up the OULAD. Any machine learning

allows for the modelling of students based on their VAK (variables algorithm cannot directly receive these data as input. Firstly, we con­
associated with learning). verted the relational tables to a dataset that can be processed by machine
learning. The input features for this study are collected from many tables
8. Experiment result and combined into a recorded dataset.
A portion of the dataset with The code_presentations 2013 J and
In this section, we will show the content of the datasets and the 2014 J is chosen for additional testing. The generated dataset rows
preprocessing steps to the data by our model to prepare the datasets with correspond to each student’s ID, and their columns to each student’s
the mapping process then discuss the experimental results. features. The missing attribute values are filled with zeros if the student
has not logged in to such activities. It does not include students who
8.1. Data collection dropped out of classes to obtain more precise and insightful results.
The data set that resulted from such interactions only contains
Studies evaluated the efficacy of the suggested strategy using the 22,593 students, their test results, and log data of their VLE connections,
Open University Learning Analytics Dataset (OU, 2017). The data set is which are represented by daily summaries of student clicks. We removed
unique because it includes stream data that have been aggregated from duplicate data and integrated the findings of the mapping phase with the
student interactions with Moodle-based learning activities in a virtual original data.
learning environment (VLE). This analyzes pupils’ behaviour as it ap­ When the attributes are examined, it is discovered that several of
pears in their activities. Twenty other activity types were also used to them are detrimental to this study because they do not have a bearing on
better categorize user engagement with the VLE. Various behaviours, our objective. Early preprocessing eliminated these unnecessary fea­
such as downloading or watching lectures, reading course materials, or tures, ultimately making the development of the predictive model
taking tests, were described by each activity category. The data set easier.
contains data on 7 courses, 32.593 students, test scores, and VLE in­ The current research aims to discover the variables that affect how
teractions (10.655.280 entries per day). It also has the results of their learners’ styles are predicted in a virtual learning environment. An
test quizzes and homework. extremely involved student will dynamically participate in the learning
environment, according to studies. Based on student participation, the
• Student demographics: includes information on the student’s age, various types of activities in OULAD are quantified. The attribute
gender, and educational level. Course Enrollment: Information on total_number_of_VLE_clicks, which tracks the learner’s overall VLE ac­
the courses that students have enrolled in, such as course codes, ti­ cess, oversees that. However, it is possible that the student will click on
tles, and start/end dates. interfaces other than the VLE throughout the session. Therefore, the
• Learning Activities: Records of how students interact with learning total of clicks alone cannot be used to determine a learner’s learning
materials, such as accessing course information, submitting assign­ preferences, so we will do a series of steps to get the best shape of the
ments, participating in conversations, and using multimedia tools. data that helps us to achieve our goal.
• Assessment Data: Details regarding how students performed on
quizzes, exams, assignments, and other assessments. 1. We will merge the learning styles with the activity type and stu­
• Clickstream Data: Student information clicks activities in the online dent_VLE_Data, which contains the number of clicks of each activity
learning environment, such as navigation patterns, time spent on and the id_site, Table 4 shows the relations between tools and ac­
different pages, and interactions with course tools and features. tivity (clicks) and their learning style orientations ((Visual, Auditory,
• Style data can be recorded as attributes in individual student profiles Kinesthetic) which identifies the number for the VLE material. After
or independent records related to student identifiers. doing that, we will select the max value of the sum of clicks that have
• Style data could comprise categorical variables that reflect students’ been done based on the id_site. We will get all the data features that
preferred learning styles. Analysing user clicks based on log data we need to start working, as shown in Fig. 2. The activity type is the
might reveal insights into students’ learning behaviours and prefer­ activity done by the students on Moodle, and the learning style is the
ences, potentially indicating their preferred learning approaches. results of his preferred study type of data to get one of the VAK
• Clickstream data can be analyzed to uncover interaction patterns students learning.
associated with learning styles. For example, students who prefer 2. Following the data merger, we encoded it since this method enables
visual learning may spend more time studying multimedia us to turn categorical variables into numerical values that can be
quickly fitted to a machine learning model. After encoding the data,
we noticed a large gap between the values in the data, so we

6
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

Table 4 8.3. Result and discussion

The relations between tools and activity (clicks) and their learning style orien­
tations (visual, auditory, kinesthetic). It is challenging to see how students will learn in different contexts.
Activity Type Learning Style Number of clicks The data preparation process, which started with raw OULAD, produced
ou_collaborate Visual 1217
a refined and condensed data set that was used to create a prediction
ou_content 16,661 model. In this study, the support vector machine, K-nearest neighbour,
questionnaire 314 Random Forest and Logistic regression were used as classification
repeat_activity 8 methods.
Resource 46,480
Both linear and non-linear solutions are supported by SVM. Addi­
URL 4398
htmlactivity Auditory 45 tionally, KNN outperforms linear regression when the signal-to-noise
Forming 9317 ratio SNR of the data is high, and it is assumed that the objects are al­
ou_elluminate 385 ways close to each other. Decision trees are less effective and accurate
ou_wiki 3018 than Random Forest. This clarifies the assortment of calculations uti­
shared_subpages Kinesthetic 13
Subpage 16,661
lized for predictive demonstration creation since it requires training
Glossary 1456 data and runs slower than decision trees. Where do we typically use
dualPane 258 Python libraries in these algorithms? Sklearn, NumPy, and Pandas are
Dataplus 512 the packages used. We ran two experiments using 90 % to 80 % of the
data for each identified student training data set. The data are tested
using the remaining 10 % and 20 % of the data, and all three assessment
metrics are used to assess them. Accuracy, recall, and outcome. To
respond to research questions, experiments are carried out.

8.3.1. Performance-based classification algorithm

This Subsection tries to address research question one, “Which
classifier offers the best performance for predicting a student’s learner
style in the VLE?” Predictive models have been developed using a va­
riety of machine-learning approaches and tests to offer a solution to this
research problem. The predicted variable in this scenario is each stu­
dent’s preferred learning method, and the characteristics were VLE ac­
tivity clicks in a VLE cycle. A 10 K-fold approval approach is used to
decide the adequacy of the categorization models. Tables 5 and 6, as
well as Figs. 4-7, include a list of evaluation metrics for the various
Fig. 2. Data after the first step (merging). models.
The random forest calculation crosses the other classification stra­
performed data scaling shown in Fig. 3 to facilitate model detection tegies in this test setup. Unlike the other models, the Random Forest
and understanding of the problem. Scaling is a technique of reducing show is a gathering model; each tree within the Random Forest produces
distances to bring them closer together. When we run machine comes about, and the course with the most votes decides the model’s
learning algorithms on the dataset, scaling the data is one of the estimate.
preprocessing procedures. Most supervised and unsupervised The accuracy of the Random Forest classification technique is 98 %.
learning methods, as we all know, base their inferences on the data The classifier has a recall of 99 % and a precision of 97 %. The classifier’s
sets that are applied to them, and many times, algorithms calculate recall performance is essential for solving the current issue of identifying
the distance between data points to draw more accurate inferences students’ learning preferences.
from the data. We have made a few important classifications. On the off chance that
the learning strategies of most understudies can be recognized at an
As we all know, many supervised and unsupervised learning methods early stage, at that point, the teaching/learning preparation will be an
rely on their conclusions on the data sets that are applied to them, and advantage. The random forest classifier is so popular that 98 % of the
frequently algorithms measure the distance between data points to overall students were predicted by their learning styles within the cur­
conclude the more accurate data. rent exploratory setting. The execution of a random forest is a marker
that the highlight under study has great prescient control. The Distri­
bution of student learner style of OULAD Dataset results in Fig. 8 the
highest percentage was 75.90 % for the visual learner style.
We applied our new algorithm in this research on our students in our
faculty on general course for the first year which all students registered
for this course (580 students).
We have a Moodle learning management system in our faculty, firstly
we prepare the material for this course to let users engaged on the

Table 5
Shows the algorithm performance using the training model of 90 and the test
model of 10.
Algorithm Accuracy Precision Recall F1-Score

Support vector machine (SVM) 84 % 81 % 99 % 89 %

K-Nearest Neighbor (KNN) 83 % 80 % 99 % 89 %
Random Forest (RFR) 98 % 97 % 99 % 99 %
Logistic regression 86 % 91 % 82 % 83 %
Fig. 3. Data after encoding and scaling steps.

7
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

Table 6
The result of algorithm performance when training mode 80 and test model 20.
Algorithm Accuracy Precision Recall F1-Score

Support vector machine (SVM) 96 % 81 % 99 % 89 %

K-Nearest Neighbor (KNN) 73 % 87 % 91 % 89 %
Random Forest (RFR) 95 % 86 % 91 % 90 %
Logistic regression 87 % 91 % 83 % 81 %

Fig. 7. Comparison between the performance of each algorithm based

on accuracy.

Fig. 4. Comparison between the performance of each algorithm.

Fig. 8. Distribution of student learner style of OULAD Dataset.

of clicks activity and recommend the material based on their learning

styles selected from the system and apply different assessment methods
Fig. 5. Comparison of the performance of each algorithm based on precision. to find the final results.
The distribution of student learner style in faculty shown in Fig. 9
that visual has the highest learner style with 60 % of the total number of
students and Table 7 shows the accuracy of different machine learning
models to measure the effectiveness of the new algorithm and accuracy.
We split the data into 90 % training and 10 % testing.

Support K-Nearest Random Logis c

vector Neighbor Forest (RFR) regression
machine (KNN)
(SVM)

Fig. 6. Comparison of the performance of each algorithm.

system with VAK learning styles, secondly, all students registered on the
site with no learning style we select the suitable style after following
their activity on the system after a period time based on logs and number Fig. 9. Distribution of student learner style in faculty.

8
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

Table 7 Table 8
Shows the algorithm performance using the training model on the faculty Result of the Prediction of the Student Assessment Method.
students. Algorithm Accuracy Precision Recall F1-
Algorithm Accuracy Precision Recall F1-Score Score

Support vector machine (SVM) 87 % 92 % 97 % 94 % Support vector machine (SVM) 96 % 97 % 96 % 97 %

K-Nearest Neighbor (KNN) 95 % 94 % 99 % 96 % K-Nearest Neighbor (KNN) 98 % 99 % 98next 98 %
Random Forest (RFR) 96 % 98 % 98 % 97 % %
Logistic regression 92 % 98 % 98 % 97 % Random Forest (RFR) 96 % 97 % 96 % 97 %
Logistic regression 95 % 96 % 97 % 97 %

9. Recommended assessment method

Each type of VAK learning style is associated with evaluation tech­

niques that students can improve. In this phase, an assessment approach
type is suggested based on the learned style of the learner. he states that
the VAK learning style assessment method types were applied, with
auditory learners favouring subjective evaluation techniques. Visual
learners embrace subjective assessment techniques instead of kines­
thetic learners who favour objective assessment techniques. Long essay
questions and case studies are two assessment methods that might be
classified as tutor-marked assessment (TMA) types but do not have a
definite right or wrong answer.
Objective assessment includes various assessment methods even if
computer-marked assessments (CMAs) of multiple-choice, true-or-false,
and other forms only need one correct answer. Based on each student’s
previously identified learning style and performance in each course, we
will choose the most appropriate assessment techniques for them shown
in Table 8. The VLE_data, Study_type, Assessment_data, and Student
Assessment Data were combined in the preprocessing step to help us
detect suitable assessment methods. Fig. 10. Result of the students’ assessment method prediction.
Performance-based classification algorithm This section seeks a so­
lution to the second RQ2 by focusing on which classifier performs the 98 %.
best in predicting a student’s appropriate evaluation technique in the We will make quick alterations to learning aspects, such as learning
VLE. Predictive models have been developed using a variety of machine- contents, evaluation procedures, and more, in our subsequent work
learning approaches and tests to offer a solution to this research prob­ while considering the student’s current learning activities.
lem. The anticipated variable in this scenario is each student’s assess­ In the future, we can propose hybrid learning techniques, and
ment technique, and the features included VILE VLE_data, Studey_type, establishing engaging and effective models for self-regulated learning
assessment_data, and student assessment data in a VLE cycle. The requires capturing the most recent technological breakthroughs and
evaluation metrics for the different models are in Fig. 10. educational research.
Content learning: The primary purpose of education is to give stu­ Investigate the use of emerging technologies such as artificial intel­
dents an in-depth understanding of the material being taught. Mastery of ligence, augmented reality, virtual reality, and immersive simulations to
knowledge means that students can understand and apply concepts develop engaging and interactive learning environments.
regardless of how they were taught. Using AI-powered adaptive learning platforms and intelligent
Full mastery of knowledge involves the ability to learn indepen­ tutoring systems to provide students with tailored feedback, recom­
dently, regardless of the learning mode. Educators should seek to mendations, and support based on their unique learning requirements
establish learning environments that promote deep comprehension, and preferences.
critical thinking, and the ability to apply knowledge in a variety of
circumstances, regardless of learning style. CRediT authorship contribution statement
Learners may choose the method of assessment as a suggestion to
their teachers, which may influence the performance rating for the final Ahmed Rashad Sayed: Methodology, Software, Data curation,
results of the test. Validation. Mohamed Helmy Khafagy: Data curation, Investigation,
Validation. Mostafa Ali: Validation, Visualization, Investigation.
10. Conclusion and future work Marwa Hussien Mohamed: Conceptualization, Methodology, Valida­
tion, Writing – original draft.
Using student models and learning-type predictions, this research
created an adaptable approach for evaluation ideas based on detected
Declaration of competing interest
learning styles. The following understudy learning exercises that use
both semantic and machine learning innovation allowed the identifi­
The authors declare that they have no known competing financial
cation of the VAK learning style of the understudy.
interests or personal relationships that could have appeared to influence
Tests were conducted utilizing the Open College Learning Analytics
the work reported in this paper.
Dataset to show the system’s capacity to recognize student learning
preferences and recommend assessment techniques. The productivity of
References
the proposal framework is surveyed using precision, accuracy, recall,
and score. According to studies, the system can determine a student’s [1] Essa SG, Celik T, Human-Hendricks NE. Personalized adaptive learning
learning preferences with an accuracy of more than 98 %, while the F1 technologies based on machine learning techniques to identify learning styles: a
test results and the evaluation recommendation accuracy both average systematic literature review. IEEE Access 2023;11:48392–409. https://doi.org/
10.1109/ACCESS.2023.3276439.

9
A. Rashad Sayed et al. Egyptian Informatics Journal 26 (2024) 100469

[2] Muñoz JLR, et al. Systematic review of adaptive learning technology for learning in [14] A.M. Ali, A. Abdelhafeez, Breaking the silence: convolutional neural networks for
higher education. Eurasian J Educ Res 2022;2022(98):221–33. https://doi.org/ sign 1 language recognition in the deaf community, doi: 10.61185/SMIJ.
10.14689/ejer.2022.98.014. [15] Alzahrani NA, Abdullah MA. Student engagement effectiveness in E-learning
[3] Hussain M, Zhu W, Zhang W, Abidi SMR. Student engagement predictions in an e- system. Biosci Biotechnol Res Commun Feb. 2019;12(1):208–18. https://doi.org/
learning system and their impact on student course assessment scores. Comput 10.21786/bbrc/12.1/24.
Intell Neurosci 2018;2018. https://doi.org/10.1155/2018/6347186. [16] A. Abdelhafeez, A.S. Aziz, N.A. Khalil, Building a sustainable social feedback loop:
[4] N. Holmes, Engaging with assessment: increasing student engagement through a machine intel-1 ligence approach for twitter opinion mining, doi: 10.61185/
continuous assessment. [Online]. Available: http://shura.shu.ac.uk/11578/. SMIJ.
[5] Amantha Kumar J, Al-Samarraie H. An investigation of novice pre-university [17] N. Alruwais, M. Zakariah, Student-engagement detection in classroom using
students’ views towards MOOCs: the case of Malaysia. Reference Librarian Apr. machine learning algorithm, Electronics (Switzerland), vol. 12, no. 3, Feb. 2023,
2019;60(2):134–47. https://doi.org/10.1080/02763877.2019.1572572. doi: 10.3390/electronics12030731.
[6] Hassan SU, Waheed H, Aljohani NR, Ali M, Ventura S, Herrera F. Virtual learning [18] Zhang Y, Yun Y, An R, Cui J, Dai H, Shang X. Educational data mining techniques
environment to predict withdrawal by leveraging deep learning. Int J Intell Syst for student performance prediction: method review and comparison analysis. Front
Aug. 2019;34(8):1935–52. https://doi.org/10.1002/int.22129. Psychol Dec. 2021;12. https://doi.org/10.3389/fpsyg.2021.698490.
[7] N.R. Aljohani, A. Fayoumi, S.U. Hassan, Predicting at-risk students using [19] Zhu X, Yang X, Huang Y, Guo Q, Zhang B. Measuring similarity and relatedness
clickstream data in the virtual learning environment, Sustainability (Switzerland), using multiple semantic relations in WordNet. Knowl Inf Syst Apr. 2020;62(4):
vol. 11, no. 24, Dec. 2019, doi: 10.3390/su11247238. 1539–69. https://doi.org/10.1007/s10115-019-01387-6.
[8] Ayar Z. The systematic review of studies about web-based learning in language [20] W. Samanthi, H. Sunesh. (2017). Relationship among students’ learning styles,
teaching context. e-Kafkas Eğitim Araştırmaları Dergisi Aug. 2021;8(2):123–36. assessment methods and students’ performances. Third International Conference
https://doi.org/10.30900/kafkasegt.957362. on Education and Distance Learning – 28th April 2017, Sri Lanka.
[9] Del Valle R, Duffy TM. Online learning: Learner characteristics and their [21] D. Soergel, WordNet. An Electronic Lexical Database New ways of learning for a
approaches to managing learning. Instr Sci Mar. 2009;37(2):129–49. https://doi. changed population of learners preparing for a changing world of work.
org/10.1007/s11251-007-9039-0. Technology-enabled knowledge support makes it possible View project DL.org-
[10] N. Alajlani, M. Crabb, I. Murray, A systematic review in understanding Coordination Action on Digital Library Interoperability, Best Practices, and
stakeholders’ role in developing adaptive learning systems, Journal of Computers Modelling Foundations View project,“ 1998. [Online]. Available: www.let.uva.nl/
in Education, Jun. 2023, doi: 10.1007/s40692-023-00283-x. ~ewn.
[11] A.M. Ali, A. Abdelhafeez, DeepHAR-Net: a novel machine intelligence approach for [22] D.S. Sharma, R. Sangal, J.D. Pawar, M.A. Tayal, M.M. Raghuwanshi, L. Malik,
human activity recognition from inertial sensors 2, doi: 10.61185/SMIJ. Word net based Method for Determining Semantic Sentence Similarity through
[12] Kotsiantis S. Supervised machine learning: a review of classification techniques. various Word Senses, NLPAI, 2014.
Informatica (slovenia) 2007;31:249–68. [23] Wu, Zhibiao & Palmer, Martha. (2002). Verbs Semantics and Lexical Selection.
[13] Ikawati Y, Al Rasyid MUH, Winarno I. Student behavior analysis to predict Proceedings of the 32nd Annual Meeting on Association for Computational
Learning styles based felder silverman model using ensemble tree method. Linguistics. 133-138. 10.3115/981732.981751.
EMITTER Int J Eng Technol Mar. 2021;9(1):92–106. https://doi.org/10.24003/
emitter.v9i1.590.

10

View publication stats