Intelligent course recommendation approach based on modified felder-silverman learning style model

Abstract

The popularity of E-learning has encouraged learners to benefit from Massive Open Online Courses (MOOC) platforms however information overload is a common challenge in order to select the appropriate courses. Recommendation systems aid learners to select relevant courses and content based on their preferences, interest and learning goals from MOOC. Various factors are considered in course recommendation systems however little research has been done in recommending courses based on learning styles despite its significance laid down by various studies. Most studies have focused on content filtering matching with learners learning style instead from a course. This research study is oriented towards learning style-based course categorization using course content delivery information data from a MOOC platform. The Felder-Silverman learning style model (FSLSM) has been used to develop a course learning style support identification model. The effectiveness of the model developed is analyzed through an experimental study performed on real-time data due to absence of a standard data set for course contents and learning style relationships. The model threshold values are finetuned through the results obtained from the experimental. The application of the course learning style identification model is demonstrated by conducting test case studies consisting of users with different learning styles and recommending courses from a MOOC platform with matching learning style.

Keywords

Course recommendation learning style online learning felder-silverman learning style model mooc E-Learning

Introduction

The evolution of technology and the Internet has provided a new era in different sectors, especially in the field of education e-learning has emerged as a new trend benefiting students from all around the globe. This combination of Internet technology with educational practices has given rise to the prominence of MOOC platforms (Bachiri and Mouncif, 2023), such as NetEase Cloud Classroom and Coursera. These platforms have significant influence over traditional educational paradigms and are rapidly reshaping the learning landscape for students. With the digitalization and widespread dissemination of educational content, the resources available on MOOC platforms are expanding exponentially, reaching exabyte levels daily. Depite the benefit is provides to learners in the form of availability of vast array of resources, it also poses the information overload problem (Ashraf et al., 2023).

Due to the varying levels of proficiency among learners and other preferences, accurately shorlisting the relevant resources becomes a increasingly intricate. This complexity can lead to confusion, arbitrary selections, and wasted time as learners tackle with navigating through the extensive array of available options. Consequently, the significance of course recommendations within MOOC platforms becomes evident (Ashraf et al., 2021). Recommendation systems have found utility in various domains, notably gaining prominence in recent years within electronic commerce (Pavlidis, 2019). With the advent of recommendation systems for product or service sales on digital platforms (Chandra et al., 2022), their application has extended to education through Learning Management Systems (LMS), aiming to enhance online learning and teaching processes. An LMS serves as web-based software designed to oversee academic courses, facilitating tutors in managing the teaching-learning dynamic by integrating students, digital resources, and content based on instructional design principles (Syed and Nair, 2018). Recommendation systems hold the potential to enhance the dissemination of academic materials to students by offering content tailored to their individual needs and learning styles (Aguilar et al., 2022). However, it’s essential to note that recommendation systems in the educational domain are not directly transposable to conventional systems, chiefly due to the unique cognitive states of students and the specialized nature of learning contents within this domain (Thongchotchat et al., 2023).

To enhance the efficacy of teaching-learning processes within virtual learning environments (VLE), the implementation of recommendation systems is imperative. These systems should undertake thorough analyses of various student variables, including learning styles, preferences, and social networks, among others. Furthermore, they should continuously recommend digital content, update information on ongoing academic courses, and fulfill other pertinent functions. Building upon these concepts, this paper proposes an intelligent autonomous architecture for a recommendation system intended for use within a VLE, aimed at optimizing course recommendations. The architecture introduces a hybrid recommendation model designed to effectively manage inputs and address specific recommendation requirements. By extracting information related to courses (such as learning objects, teaching strategies, and instructional designs) and students (including their learning styles and socioeconomic backgrounds), as well as details about the courses themselves (such as instructional designs), the proposed architecture facilitates intelligent and autonomous recommendations (Monsalve-Pulido et al., 2024).

The paper is structured as follows: In section 2, the related work previously done in the field is discussed. In section 3, the proposed methodology for devising the course learning style identification model and course recommendation approach is discussed. Section 4 is utilized for the discussion of experimentation and result analysis. In section 5, the conclusion is discussed along with the scope of future works in the field.

Related work

This section offers a comprehensive detail of the current review papers in the domain following a thorough analysis of educational expert recommender systems with a focus on learning style-based recommendations.

Recommender systems in E-learning

The advancement of recommender systems in e-learning represents a dynamic and extensively researched field. A plethora of review papers focusing on Recommender Systems (RS) in e-learning have been documented in the literature. These RSs are designed to propose courses to students based on a multitude of criteria, ranging from individual interests and academic fields to performance metrics and prior user ratings. The reviews encompass a comprehensive examination of the primary paradigms within recommender systems, encompassing both explicit and implicit feedback mechanisms. Furthermore, they explore various methodologies employed in the design of recommender systems aimed at augmenting the learning experience (Kulkarni et al., 2020).

Khanal et al. (2020)have conducted a comprehensive review of recommendation systems within the context of e-learning, culminating in the development of a taxonomy that encompasses essential components essential for crafting an effective recommendation system. Their study underscores the indispensability of machine learning techniques, algorithms, datasets, evaluation metrics, valuation criteria, and output mechanisms in the construction of such systems. This paper stands as a noteworthy contribution to the field, offering a vital overview of the existing research landscape while shedding light on persisting challenges.

Dhananjaya et al. (2024) have contributed a survey paper aimed at highlighting the transformative benefits stemming from the integration of adaptive technologies into personalized recommendation systems within the realm of education. Their work emphasizes the development of a comprehensive Personalized Education Recommendation System, which leverages advanced technologies such as NeuroSky EEG and AI Chatbots. This integration holds the promise of revolutionizing traditional learning paradigms by fostering adaptive, engaging, and effective learning experiences tailored to individual student needs. Features like Alter Ego and Twin Technology are poised to support personalized learning paths by dynamically adjusting content and resources to accommodate individual student requirements and learning styles. Additionally, the paper outlines a generic structure of recommendation systems in the e-learning domain as depicted in Figure 1.

Figure 1.

General framework of recommendation process.

Learning style-based recommender systems

The recommendation systems for learning styles in research literature are focused on matching the content from a course. Chen et al. (2020) introduced a novel model founded on modified Collaborative Filtering (CF), termed Adaptive Recommendation based on Online Learning Style (AROLS). This model records learning styles through eight distinct features. Recommendations for sets of learning materials are generated using a combination of CF and association rule mining techniques, which cluster similar learners to provide tailored suggestions.

Baidada et al. (2019) proposed a hybrid recommendation system that combines both Content-Based (CB) andCF techniques to address the cold start problem. This system focuses on offering teaching aids, learning objects, and content complexity levels tailored to a selected course based on the user’s academic skills and learning style profile. While the recommender system is implemented in a hybrid e-learning and face-to-face learning environment, its efficacy in MOOCs remains unexplored. On the other hand, (Bhaskaran and Santhi, 2019) have utilized a hybrid strategy integrating the firefly algorithm and the k-means algorithm to cluster learners based on their learning styles. They mined the preferences of learners appearing in frequent sequences using the AprioriAll algorithm and customized recommendations using the trust-based weighted mean approach. Their framework exhibited performance enhancements in terms of both accuracy and speed. Nonetheless, further exploration of other hybrid optimization algorithms is warranted to enhance accuracy further.

Nafea et al. (2019) have employed a combination of Content-Based Recommender Systems (CBRS) and Collaborative Filtering Recommender Systems (CFRS) to enhance prediction accuracy and recommend the most suitable course learning objects (LOs). This approach takes into account student learning styles, LO profiles, and students’ ratings of LOs to mitigate the cold-start problem. In a similar vein, (Yadav and Sohal, 2017) have leveraged the Genetic Algorithm (GA) to personalize learning content within an e-learning system. They considered various factors including the learner’s knowledge level, learning style, interactivity preferences, and complexity of learning objects. However, these studies primarily focused on providing personalized learning objects within a course without delving into the analysis of the course structure.

Aeiad and Mezaine (2019) have introduced APELS, an E-learning system that places a strong emphasis on personalization and adaptability. APELS tailors learning experiences by taking into account individual learner attributes such as background, needs, and learning styles. Through a computer science case study aligned with standard curricula, APELS generates personalized learning materials. It utilizes ontology and rule-based methods to sift through online resources and evaluate content against learning outcomes. Empirical evaluation validates APELS’ capability to generate high-quality, personalized learning materials, highlighting its potential to enhance E-learning by catering to specific learner attributes.

The integration of Web 2.0 technologies, particularly social media, has brought about a significant transformation in collaborative education. However, prevailing LMS often lack personalized support tailored to individual learner attributes, such as knowledge level estimation and learning style. In response to this gap, Abri et al. (2020) have proposed a framework aimed at enhancing virtual learning environments by integrating social media tools for seamless collaboration and adaptive learning. This framework leverages content generated during collaboration to personalize learning experiences based on discussed concepts and learner characteristics. Initial evaluations indicate promising results, highlighting the potential to optimize student engagement and enhance personalized learning in e-learning environments. Ongoing research in this realm holds promise for fostering improved student collaboration, participation, and understanding of individual student needs, thus facilitating effective personalized learning experiences.

Numerous researchers have delved into online course design for various objectives. For instance, Martin et al. (2021) have devised an instrument to pinpoint online course design elements aimed at sustaining student engagement and motivation. However, these investigations primarily concentrated on the impact of diverse course components on augmenting student learning outcomes, rather than scrutinizing course design to identify specific learning style groups. The suitability of a course for a particular learning style hinges on whether its materials resonate with that specific learning type. Despite the abundance of courses available in LMSs, scant attention has been paid to how effectively these courses cater to students’ learning styles. Existing research on e-learning course design employing learning styles predominantly revolves around evaluating learning objects based on learning style and proposes personalizing learning objects within a course by adjusting their order and placement (Laksitowening et al., 2016). However, few studies have delved into the mechanism for assessing course learning style while neglecting critical factors like time-based activities in e-courses (El-Bishouty et al., 2019).

To bridge this research gap, this paper introduces a learning style identification model for online courses, enabling the early assessment of compatibility with potential learning style categories. This model serves to aid students and stakeholders in gauging an online course’s suitability for their preferred learning style, thereby augmenting the likelihood of achieving successful learning outcomes. Importantly, this model will constitute a foundational component of a course recommendation approach aimed at filtering courses that align with learners’ learning styles from MOOCs.

Methodology

The methodology section outlines the systematic approach taken to develop a course recommender system based on learning styles as depicted in Figure 2. The process involves several key steps. The first step involves selecting course data that influences learning styles. The relationship between course data and learning styles is intricate and influenced by various factors affecting how learners interact with course content. Research studies have been pivotal in establishing this relationship, focusing on different types of learners such as visual, auditory, and kinesthetic learners. The second step is to create a model that quantifies the type of content delivery data and sets threshold boundaries based on the FSLSM. The model developed aims to predict the level of support a course provides for a specific learning style. It is theoretically grounded and requires empirical validation through case studies. The hypotheses regarding the threshold boundaries between different learning styles are tested and refined based on the data collected during the case study. After testing and validation of the model, the course recommender system is developed that is used to recommend courses tailored to individual learning styles.

Figure 2.

Research methodology.

LS and course data relationship

In academic literature, it has been well-documented that the level of support for different learning styles (LS) is dependent upon the instructional media (IM) and learning objects (LO) used to deliver the course contents. This distinction is highlighted through the categorization of course content data into two main groups: teacher-centered and learner-centered, as illustrated in Table 1.

Table 1.

Course IM/LO classification.

Teacher-centered	Learner-centered
E-lectures	Reading material
Recitations	Assignments/Exercises
Tutorials	Quiz/Exam
Labs	Projects

The teacher-centered group encompasses course materials that are presented or overseen by the instructor, such as E-Lectures, Recitations, Tutorials, and Labs, each with predetermined time allocations. These components are designed to be instructor-led and structured within specific time frames.

On the other hand, the learner-centered group focuses on student-driven instruction, where learning occurs through reading materials, projects, assignments/exercises, exams/quizzes, and group discussions. In this approach, students have the autonomy to engage with the materials at their own pace without strict time constraints. Consequently, the course content within this group is not bound by time limitations, allowing learners to assimilate the covered material through various learning objects like course projects, exercises/assignments, readings, and quizzes/exams (Study.com, 2022).

IM/LO relationship with felder-silverman learning style model

The IM/LO listed in Table 1, provides support for different types of learners. This section explains the IM and LO while establishing their connection with the LS categories of FSLSM. The FSLSM provides a comprehensive framework that categorizes individual learning preferences into four dimensions: Active-Reflective, Sensing-Intuitive, Visual-Verbal, and Sequential-Global. The model is built upon the notion of cognitive diversity and the model identifies four dimensions on which the learners vary. This influential model has significantly contributed to pedagogical strategies, fostering an inclusive and adaptive approach to education that acknowledges the multifaceted nature of how individuals process and internalize information. The instructional methods and learning objects can be aligned with Felder-Silverman’s LS model, which encompasses four dimensions as given in Table 2.

Table 2.

Felder-silverman LS model.

Dim 1	Active learners	They prefer hands-on experiences and interaction and benefit from experiments, discussions, and active projects
Dim 1	Reflective learners	They contemplate by thinking critically, take time to think deeply, and utilize writing and introspection to consolidate learning
Dim 2	Sensing learners	They prefer practical, hands-on, step-by-step approaches, and benefit from structured plans, repetition, and practical application
Dim 2	Intuitive learners	They embrace abstract concepts and theories, explore connections between ideas, and focus on big-picture perspectives independently
Dim 3	Visual learners	They grasp concepts through images and visual aids, engage well with charts, graphs, and visual presentations, and benefit from visual aids like diagrams and mind maps
Dim 3	Verbal learners	They learn through written and spoken language, flourish in discussions, lectures, and reading, and utilize reading, note-taking, debates, and discussions
Dim 4	Global learners	They understand by grasping the overall context, focus on connections and relationships, and enjoy exploring possibilities and patterns through concept mapping, problem-solving, and brainstorming
Dim 4	Sequential learners	They learn through a logical sequence of information, excel at processing specific facts and details, prefer organized and structured learning experiences, and utilize outlining step-by-step approaches, and logical sequences

By considering FSLSM following relationship can be established.

E-Lectures

E-lectures offer benefits for various learning styles. Reflective learners benefit from the ability to control the pace, pause, and review content for thoughtful reflection. Sensing learners are supported through visual aids, real-life examples, and interactive elements like quizzes. Visual learners are accommodated with slides, images, and videos, aiding comprehension and retention. Sequential learners appreciate structured content that follows a predefined path, with clear outlines or summaries aiding the organization and synthesis of the material.

Recitation

Recitation sessions offer various benefits tailored to different learning styles. For active learners, group discussions, role-playing, and hands-on experiments provide opportunities for engagement and collaboration. Intuitive learners benefit from activities such as concept mapping, case studies, and open-ended questions, which stimulate critical thinking and encourage exploration. Verbal learners are supported through interactive discussions, presentations, and written tasks, allowing them to express themselves verbally and enhance their communication skills. Overall, recitations cater to diverse learning styles by providing a range of activities that accommodate different preferences and strengths.

Tutorials

Tutorials offer support for reflective learners through self-paced learning, pause and review features, and reflection prompts. Learners can proceed at their own pace, pause to review content, and engage in reflective activities to deepen understanding. Additionally, tutorials cater to intuitive learners by providing conceptual explanations, visual representations, and real-world examples. Clear explanations and visual aids help intuitive learners grasp complex concepts, while real-world examples illustrate practical applications, fostering deeper comprehension and connection-making.

Labs

Labs support active learners with hands-on experiments and collaboration, allowing them to apply theoretical knowledge practically and engage in problem-solving. They cater to sensing learners through concrete experiences, data analysis, and practical examples, satisfying their preference for tangible learning. Visual demonstrations and representations in labs accommodate visual learners, aiding their comprehension through observation and visualization of concepts.

Individual assignment

Individual assignments support reflective learners through open-ended tasks, written reflections, and self-pacing. For sensing learners, assignments offer concrete examples, clear instructions, and hands-on application, engaging their attention to detail. Verbal learners benefit from written assignments and presentations, enhancing communication skills. Sequential learners find support in structured tasks and clear guidelines, facilitating methodical progression through assignments.

Projects

Projects support active learners through hands-on engagement and teamwork, fostering problem-solving and decision-making skills. For intuitive learners, projects offer open-ended exploration and creative problem-solving opportunities, connecting theoretical concepts to practical applications. Additionally, projects accommodate global learners by emphasizing big-picture perspectives, interdisciplinary approaches, and real-world connections, allowing them to understand the broader implications of their work.

Readings

Reading supports reflective learners through extended reading time, note-taking, and journaling activities, allowing them to engage deeply with the material and articulate their understanding. For intuitive learners, reading materials provide conceptual frameworks, analogies, and real-world examples to grasp complex ideas and make connections. Visual aids such as infographics accommodate visual learners, while clear and concise writing supports verbal learners. Sequential learners benefit from the clear organization and step-by-step instructions found in reading materials.

Quizzes/exams

Quizzes/exams benefit reflective learners by enabling self-assessment, review, and feedback. They support intuitive learners through application-oriented and open-ended questions that encourage critical thinking and creative problem-solving. Additionally, quizzes/exams assess an intuitive understanding of underlying principles, fostering connections across topics and the ability to grasp the bigger picture.

The summarized form of mapping of IM/LO with LS has been given in Table 3.

Table 3.

Matching IM/LO with LS.

IM/LO	LS
	Dim 1 perception		Dim 2 entry channel		Dim 3 processing		Dim 4 understanding
	Active	Reflective	Sensing	Intuitive	Visual	Verbal	Sequential	Global
E-lectures		√	✓		√		√
Recitations	√			✓		√
Tutorials		√		√
Labs	√		√		√
Ind. Assignments		√	√			√	√
Projects	√			√				√
Readings		√		√	√	√	√
Quiz/Exams		√		√
Presentations	√			√	√	√	√

Course LS identification model

The course LS identification model is designed to identify the course LS for the four dimensions of FSLSMThis model evaluates the course’s alignment with both ends of each LS dimension by analyzing the IM and LO data. The data from the two distinct groups, teacher-centered and learner-centered, are processed independently. Quantitatively, LS support is determined by summing the time spent on teacher-centered activities for both ends of each dimension. This cumulative value is then normalized on a scale of 10 for each LS dimension. Similarly, learner-centered activities are tallied and aggregated, with different activities carrying varying weights; for instance, project assignments may hold more significance than individual tasks. The formula below illustrates the calculation method to ascertain the support level for each of the four FSLSM dimensions, aiding in the identification of the course’s LS alignment:

\begin{array}{l} f o r \dim i w h e r e i = 1 \to 4 \\ L S S u p p o r t f o r P o l e 1, \dim i = \frac{t i m e (T C) f o r P o l e 1}{T o t a l t i m e f o r T C, \dim i} + \frac{s u m o f (L C) o b j e c t s f o r P o l e 1}{T o t a l L C o b j e c t s, \dim i} \\ L S S u p p o r t f o r P o l e 2, \dim i = \frac{t i m e (T C) f o r P o l e 2}{T o t a l t i m e f o r T C, \dim i} + \frac{s u m o f (L C) o b j e c t s f o r P o l e 2}{T o t a l L C o b j e c t s, \dim i} \end{array}

The difference in the results of both poles is used to categorize the Course LS. Figure 3 presents a chart that is used to interpret the results. The chart categorizes the LS of each dimension in five categories that is, strong pole 1, medium pole 1, fairly balanced, medium pole 2, and strong pole 2. This approach is inspired by the Felder Silverman method, where the poles are distributed in the interval of −11 to +11. Four threshold boundaries divide the LS into five categories. Initially, the values of threshold boundaries areset at configurable intervals.

Figure 3.

Course Ls intensity level interpretation chart.

Testing and optimization of model

Due to the lack of a standard data set for Course LS versus the IM/LO, the Course Identification Model is validated through real-world data of online learning in a local university. The model has four threshold boundaries that split the course classification into strong pole 1, medium pole 1, fairly balanced, medium pole 2, and strong pole 2. This makes it a multiple threshold classification model as it has multiple decision boundaries in which the goal is to classify input samples into multiple classes based on different decision boundaries. To set the threshold boundaries (T1,T2,T3,T4), the initially determined values as shown in Figure 3 are required to be tested and optimized for better results. In this regard, a fitness function is defined in which the experimental data is used for the optimization of the threshold values using GA and Surrogate Optimization (SO) algorithms. The model accuracy is adjusted through two evaluation metrics, the Mean Absolute Percentage Error (MAPE) and the Root Mean Square Deviation (RMSD). The process flow of the validation is shown in Figure 4.

Figure 4.

Course Ls identification model validation process flow chart.

The data of student LS and student feedback is compared with the Course LS obtained through the Course LS Identification Model. The threshold intervals and the weight of course data are adjusted in the model to bring the results closer. Once the results are sufficiently closed the iteration is terminated. The fitness function is designed as below:

F i t n e s s F u n c t i o n = Min (\sum_{i = 1}^{n} | {E D}_{P} - {E D}_{A} |)

(1)

In the fitness function, the ED_pis the Euclidean distance between the Course LS and the student LS while ED_A is the corresponding Euclidean distance interpreted from the feedback score of the student as shown in Table 4.

Table 4.

Satisfaction score relationship with ED.

Satisfaction score	10	9	8	7	6	5	4	3	2	1	0
Euclidean distance	0	0.8	1.6	2.4	3.2	4.0	4.8	5.6	6.4	7.2	8.0

The value of ED_A is constant while ED_pis a variable in which the course LS is changed by adjusting the threshold values T1, T2, T3, T4 that are taken as design variables. The design variables that are the threshold values have bounded, where T2>T1 & T2<10, T4<T3 & T4 >-10. The values of T1 and T3 enclose the class that shows a fairly balanced class. The threshold T2 separates the strong pole one and moderate pole one category. Similarly, the threshold T4 classifies Pole two into strong and moderate poles. Evolutionary algorithms find valuable application in multi-class optimization, a domain characterized by intricate trade-offs and conflicting objectives. Genetic Algorithm has been applied for multi-class optimization in various research studies (Khanse et al., 2020; Varniab et al., 2019). GA is a powerful algorithm that can be designed to achieve an accuracy as high as 100% in applications where its stochastic nature would not be a setback (Khanse et al., 2020). The surrogate method has been used to optimize the decision thresholds in research studies and is found to reach optimization values more efficiently (Pellegrini and Masquelier, 2021). Research studies show that the Surrogate MATLAB function has been used for optimization that achieves similar results as GA but in a shorter duration (Mahmood and Ismail, 2021).

Course recommender based on LS

To recommend matching courses with the learner’s LS, a mechanism is designed, the process flow of which is shown in Figure 5. Initially, the course data is downloaded and the LS of all courses are obtained using the LS Identification Model. This step provides a database of courses along with their LS. To group similar courses in groups k-means clustering is used to group courses with similar LS. The recommender model requires the LS of the student that is obtained through the FSLSM questionnaire. The learner’s LS and the closest cluster matching with the LS are selected. Finally, the Euclidean Distance between the LS of the courses within the selected group and the student is calculated and the closest match is recommended to the student.

Figure 5.

Flow chart of course recommendation model.

Results

Results have been segregated on basis of course learning style model and course recommendation in subsequent section.

Course learning style model results and optimization

Due to the lack of a standard data set, thecourse LS identification model is implemented on a case study conducted in a university during the online learning phase during the COVID-19 era. Two sections were selected that were enrolled in five subjects. The course content delivery data was obtained from instructors for the 10 courses that are listed in Table 5.

Table 5.

Course data from instructors.

Sr	Course	E-Lectures		Recitations		Labs		Assign	Project	Reading	Exam
Sr	Course	No.	Hrs	No.	Hrs	No.	Hrs	Assign	Project	Reading	Exam
1	Sub 1	4	2.0	0	0	4	3	2	1	1	0
2	Sub 2	8	1.5	4	2	0	0	2	0	1	1
3	Sub 3	4	2.0	0	0	4	3	1	1	1	0
4	Sub 4	8	1.5	8	1	0	0	2	0	4	0
5	Sub 5	4	2.0	0	0	4	3	1	0	1	2
6	Sub 6	8	1.5	8	2	0	0	0	0	3	2
7	Sub 7	8	1.0	0	0	6	3	1	1	0	0
8	Sub 8	8	1.5	8	1	0	0	2	0	2	1
9	Sub 9	8	2.0	4	1	0	0	4	0	0	2
10	Sub 10	4	2.0	0	0	4	3	1	1	1	1

The LS of the courses is obtained through the methodology presented in Section 3.2. The results of the LS of courses areshown in Table 6.

Table 6.

Course LS Style using initial Threshold Values.

Course	Learning style				Initial thresholds
Course	Dim 1	Dim 2	Dim 3	Dim 4	LS1	LS2	LS3	LS4
Sub 1	1.8	2.8	2.5	4.2	FB	M.Sen	M.Vis	M.Seq
Sub 2	−6.0	0.0	−1.5	10.0	M.Ref	FB	FB	S.Seq
Sub 3	2.7	1.7	3.3	2.0	M.Act	FB	M.Vis	FB
Sub 4	−6.0	−1.7	0.0	10.0	M.Ref	FB	FB	S.Seq
Sub 5	−4.0	2.5	3.3	10.0	M.Ref	M.Sen	M.Vis	S.Seq
Sub 6	−4.3	−5.0	−0.7	5.0	M.Ref	M.Int	FB	M.Seq
Sub 7	6.2	2.0	0.0	2.0	S.Act	FB	FB	FB
Sub 8	−6.0	−1.0	−0.7	10.0	M.Ref	FB	FB	S.Seq
Sub 9	−8.0	1.7	−2.0	10.0	S.Ref	FB	FB	S.Seq
Sub 10	1.7	1.4	3.3	2.0	FB	FB	M.Vis	FB

The results of the course LS identification module can be fine-tuned through the satisfaction score data obtained from the students. The threshold values have been initially chosen giving equal intervals to each level in the four dimensions of LS. Considering the relationship explained in Table 4 between the student satisfaction score and ED. between student LS and course LS the threshold values are optimized through the fitness function. The design variables are the threshold values T1, T2, T3 and T4 whose bounded limit values are given in Table 7.

Table 7.

Design variables limits.

Variables	Lower limit	Upper limit
T₁	0	5
T₂	3	8
T₃	−5	0
T₄	−8	−3

The results of the GA and SO algorithms show that the Surrogate Algorithm performs marginally better than GA, and its results are used to modify the threshold values. Figures 6 and 7 shows threshold optimization using GA and SO respectively. Figure 8 shows the modified threshold result diagram that is used to obtain the Course LS support levels.

Figure 6.

Threshold Optimization using GA.

Figure 7.

Threshold Optimization using SO.

Figure 8.

Optimized course Ls intensity level interpretation chart.

The comparison of results obtained is summarized in Table 8.

Table 8.

Optimization results.

	Genetic algorithm	Surrogate
Fitness value	0.69	0.65
T₁	3.3	3.30
T₂	8.0	7.56
T₃	−3.6	−3.60
T₄	−7.0	−6.0

The updated threshold values are used to interpret the quantitative results of the course LS values. The modified results for the course LS support are provided in Table 9.

Table 9.

Course LS Style using optimized Threshold Values.

Course	LS result				Optimized thresholds
Course	LS1	LS2	LS3	LS4	LS1	LS2	LS3	LS4
Sub 1	1.8	2.8	2.5	4.2	FB	FB	FB	M.Seq
Sub 2	−6.0	0.0	−1.5	10.0	M.Ref	FB	FB	S.Seq
Sub 3	2.7	1.7	3.3	2.0	FB	FB	M.Vis	FB
Sub 4	−6.0	−1.7	0.0	10.0	M.Ref	FB	FB	S.Seq
Sub 5	−4.0	2.5	3.3	10.0	M.Ref	FB	M.Vis	S.Seq
Sub 6	−4.3	−5.0	−0.7	5.0	M.Ref	M.Int	FB	M.Seq
Sub 7	6.2	2.0	0.0	2.0	M.Act	FB	FB	FB
Sub 8	−6.0	−1.0	−0.7	10.0	M.Ref	FB	FB	S.Seq
Sub 9	−8.0	1.7	−2.0	10.0	S.Ref	FB	FB	S.Seq
Sub 10	1.7	1.4	3.3	2.0	FB	FB	M.Vis	FB

A comparison between the initial and optimized models reveals that the initially estimated threshold values yielded satisfactory outcomes. Although optimization led to some enhancements, they were marginal. Table 10 presents the total average of MAPE values obtained from both initial and optimized models. The findings indicate a 2% enhancement in MAPE and an 11% improvement in RMSD values.

Table 10.

Results of evaluation metrics for course LS support model.

	RMSD	MAPE (%)
Initial design	1.2	17.8
Optimized design	1.07	16.2
% Improvement	11 %	2

Course recommender results

This section illustrates the functionality of the proposed model in the context of courses offered on a MOOC platform. The LS Identification model is utilized to gather learning style (LS) data for the courses. Subsequently, tailored course recommendations are provided for hypothetical users with different learning styles.

MOOC course LS results

The data of IM and LO were obtained through manual web scrapping of courses of computer science from MOOC platforms. The data of 129 courses is used in the course LS identification model to obtain a database of courses along with their support for particular. The support for the FSLSM in its four dimensions is obtained and presented as a comparative percentage in Table 11.

Table 11.

MOOC course LS results.

FSLSM Dim 1	S.Act	M.Act	FB	M.Ref	S.Ref
FSLSM Dim 1	0 %	3.1 %	38.8 %	21.7 %	36.4 %
FSLSM dim 2	S. Sen	M. Sen	FB	M.Int	S.Int
FSLSM dim 2	7 %	13.2 %	35.7 %	7.0 %	37.2 %
FSLSM dim 3	S. Vis	M. Vis	FB	M. Ver	S. Ver
FSLSM dim 3	6.2 %	40.3 %	52.7 %	0.8 %	0 %
FSLSM dim 4	S. Seq	M. Se	FB	M. Gbl	S. Gbl
FSLSM dim 4	66.7 %	22.5 %	10.9 %	0 %	0 %

In the first dimension, courses favoring the Reflective pole hold a dominant position, comprising 58% of the total. The rest tend to a balanced distribution, with only a minority leaning towards the active pole, albeit at a moderate level. Moving to the second dimension, courses aligned with the Intuitive pole constitute the majority, accounting for 44% of the total. Another 35.7% exhibit a fairly balanced distribution, while the lowest percentage, 20%, corresponds to courses supporting the Sensing LS pole. In the third dimension, a significant portion of courses falls into the fairly balanced category (52.7%), followed closely by those endorsing Visual LS (46.5%). Only a minor fraction, 0.8%, aligns with Verbal LS, which also falls within the moderate level. Lastly, in the fourth dimension, the majority of courses support Sequential LS, comprising 89.2% of the total, while the remainder falls into the fairly balanced category.

Clustering of courses

The courses divided into clusters make the search algorithm efficient for course recommendations to students. Grouping similar courses based on LS clustering has been done using k-means. The value of k = 7, based on the silhouette values after the value of k = 7 the silhouette value did not increase significantly further as shown in Figure 9. Furthermore, some values are in the negative axis as well for k = 8. To obtain optimized clusters for k = 7, a total 50 number of replicates having a maximum of 100 iterations each have been selected. In each replicate starting point values are varied to compute the cluster groups. The value of the sum of distances shows the cluster optimization value. Figure 10 shows the variation of the sum of distance with replicates in which the minimum value comes out to be 78.0404.

Figure 9.

Silhouette Value w. r.t k.

Figure 10.

Sum of distances with replication.

Test cases results

To demonstrate the recommender system designed, hypothetical test cases of students with different LS are selected and the result of the recommended course is analyzed. By assuming students with different combinations of LSs, the proposed framework was able to simulate and assess the efficacy of the recommendation algorithm. The courses are classified into seven cluster groups based on course LS. The student LS is obtained from FSLSM, and the closest cluster group of MOOC courses is identified in the first step. The closest matching course within the cluster group with the LS of the student is computed through the Euclidean Distance formula.

Case 1

The first test case is selected to demonstrate the absolute match between the student LS and course LS. The student has the LS as presented in Table 12.

The closest matching cluster group of courses comes out to be the Fifth cluster as shown in Table 13 having the least Euclidean distance between the Test Case 1 LS and the centroid of the cluster group.

In cluster 5, there are 27 courses and the result shows that there are two courses that exactly match (i.e. Euclidean distance = 0) with the LS of the student as shown in Table 14.

Case 2: The second test case is a student with LS who has a maximum possible mismatch with the courses listed in the database. The detail of student LS is given in Table 15.

The third cluster group is the nearest match obtained through the Euclidean distance between the cluster LS centroid values and the test case centroid as shown in the following Table 16.

Out of 13 courses present in Cluster 3, the closest matching course recommended has an Euclidean distance of 4.36, as shown in Table 17.

Table 12.

Student LS for test case 1.

FSLSM	Dim 1	Dim 2	Dim 3	Dim 4
Test case 1	FB (3)	S.Int (1)	FB (3)	S.Seq (5)

Table 13.

Results for Course and Student LS matching for Test Case 1.

Cluster	Centroid values				ED from TC#1 LS
1	1.125	3.125	4	5	3.005
2	1.5	1	3.25	3	2.512
3	3.3077	3.6923	4.2308	4.1538	3.094
4	2.5926	3.2593	2.963	5	2.296
5	2.8077	1.1154	3.1154	4	1.031
6	2.5	5	3.125	5	4.033
7	1.0435	1.2609	3.9565	4.8261	2.200

Bold values show the minimum Eucledian Distance between the course and student LS for the selection of recommended course.

Table 14.

Recommended courses for test case 1.

	Student LS	FB	S.Int	FB	S.Seq
Recommended courses						ED	% match
1	Artificial intelligence	FB	S.Int	FB	S.Seq	0	100 %
2	Cryptography 1	FB	FB	FB	S.Seq	0

Table 15.

Student LS for test case 2.

FSLSM	Dim 1	Dim 2	Dim 3	Dim 4
Test case 2	S.Act	S.Sen	S.Ver	S.Gbl

Table 16.

Results for Course and Student LS matching for Test Case 2.

Cluster	Centroid values				ED from TC#2 LS
1	1.125	3.125	4	5	6.598
2	1.5	1	3.25	3	6.108
3	3.3077	3.6923	4.2308	4.1538	4.996
4	2.5926	3.2593	2.963	5	5.355
5	2.8077	1.1154	3.1154	4	5.777
6	2.5	5	3.125	5	5.174
7	1.0435	1.2609	3.9565	4.8261	7.281

Bold values show the minimum Eucledian Distance between the course and student LS for the selection of recommended course.

Table 17.

Recommended courses for test case 2.

	Student LS	S.Act	S.Sen	S.Ver	S.Gbl
Recommended courses						ED	% match
1	Advanced circuit technique	M.Act	S.Sen	M.Vis	M.Seq	4.36	45.5

Case 3

The third case is selected to demonstrate a student whose LS is selectedrandomly,and its goal is to obtain 4 number of courses having the closest match with his LS. The student LS is depicted in Table 18.

Table 18.

Student LS for test case 3.

FSLSM	Dim 1	Dim 2	Dim 3	Dim 4
Test case 3	M. Ref	FB	M. Vis	FB

The nearest match to the student LS comes out to be the third cluster group obtained through the Euclidean distance between the cluster LS centroid values and the test case centroid as shown in Table 19.

Table 19.

Results for Course and Student LS matching for Test Case 3.

Cluster	Centroid values				ED from TC#2 LS
1	1.125	3.125	4	5	2.187
2	1.5	1	3.25	3	2.194
3	3.3077	3.6923	4.2308	4.1538	1.890
4	2.5926	3.2593	2.963	5	2.344
5	2.8077	1.1154	3.1154	4	2.447
6	2.5	5	3.125	5	3.003
7	1.0435	1.2609	3.9565	4.8261	2.697

Bold values shows the minimum Eucledian Distance between the course and student LS for the selection of recommended course.

The top four closest matching courses out of the 13 courses present in Cluster three are recommended as shown in following Table 20.

Table 20.

Recommended courses for test case 1.

	Student LS	M.Ref	FB	M.Vis	FB
Recommended courses						ED	% match
1	Distributed computer systems engineering	FB	FB	M.Vis	M.Seq	1.414	82.3
2	Practicing programming in C	FB	FB	M.Vis	M.Seq	1.414	82.3
3	Introductory digital systems laboratory	FB	FB	M.Vis	M.Seq	1.414	82.3
4	Designing and executing Information security strategies	FB	FB	M.Vis	M.Seq	1.414	82.3

Conclusion

This paper introduces a course recommendation system tailored to student’s learning styles to assist them in selecting courses from MOOC platforms. At the core of this system is the Course Learning Style Identification Model (CLSIM), which examines and quantifies the relationship between course content and instructional methods to determine their alignment with various learning styles, utilizing the FSLSM methodology. The CLSIM underwent testing, refining threshold values through empirical data from real-time case studies to effectively distinguish between different learning styles. Demonstrating its functionality, the recommender model framework was applied to MOOC data, yielding results for students with diverse learning styles.

While this research presents an initial model, its reliance on a limited dataset is acknowledged as a constraint. Future endeavors could significantly enhance the model’s utility by exploring larger datasets or devising methods to collect data from diverse MOOC platforms, thereby improving its ability to offer tailored course recommendations and enrich the learning experience for a broader range of students. The current recommendation model focuses solely on matching course learning styles with student learning styles. However, future iterations could develop a multi-objective version that incorporates factors such as learning style compatibility, grade prediction, course complexity, specialization, and student feedback.

The research has developed a course recommendation framework specifically targeting the Computer Science department, identifying course learning styles and aligning them with students. Future efforts will broaden the framework’s scope to encompass other academic disciplines, assessing its adaptability and effectiveness across various fields. This expansion aims to provide personalized course recommendations tailored to students’ diverse areas of study.

Footnotes

Acknowledgements

I am grateful to my co-authors, who helped me organize the concepts and proofread this manuscript.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Shams ul Arfeen Laghari

Author Biographies

Erum Ashraf is working as assistant professor at Bahria University, Islamabad, Pakistan. She has earned Ph.D in Artificial Intelligence in the area of Digital Education from National Advanced IPv6 Centre, Universiti Sains Malaysia. Her research interests include Software Testing, IoT, Industry 4.0, Cloud Computing and Artificial Intelligence.

Shams ul Arfeen Laghari received the B.Sc. (Hons.) and M.Sc. degrees in computer science from the University of Sindh, Jamshoro, Pakistan, and the M.S. degree in computer science from PAF-KIET, Karachi, Pakistan. He has completed the Ph.D in network security from National Advanced IPv6 Centre, Universiti Sains Malaysia. His research interests include cybersecurity, Industry 4.0, distributed systems, cloud computing, and mobile cloud computing.

Selvakumar Manickam is currently the Director of the National Advanced IPv6 Centre and an Associate Professor specializing in cybersecurity, the Internet of Things, Industry 4.0, cloud computing, big data, and machine learning. Previously, he was with Intel Corporation and a few start-ups working in related areas before moving to academia. He has authored or coauthored more than 220 papers in journals, conference proceedings, and book reviews. He is involved in various organizations and forums locally and globally. While building his profile academically, he is still very involved in industrial projects involving industrial communication protocol, robotic process automation, machine learning, and data analytics using open-source platforms. He also has experience in building the IoT, embedded, server, mobile, and web-based applications.

Khurrum Mahmood is working as a consultant in a design and development firm that provides support to public sector and private sector industry. His area of expertise is Multi-disciplinary design optimization and has publications in this area in reputable journals and conferences. His qualifications includes M.Sc in Industrial and Manufacturing Engineering from UET Taxila Pakistan, M.Sc in Aerospace Engineering from USM Malaysia.

Sehrish Abrejo received her Bachelor’s of Computer Science degree from Isra University, Hyderabad. She continued her further studies in the same discipline and received MSCS, M.Phil and PhD degrees from same institute. She is working as Assistant Professor in the department of Computer Science at Isra University, Hyderabad, Pakistan. Her research interests include Mobile ad hoc Networks, Artificial Intelligence and Natural Language Processing.

Shankar Karuppayah received the B.Sc. degree (Hons.) in computer science from Universiti Sains Malaysia, in 2009, the M.Sc. degree in software systems engineering from the King Mongkut’s University of Technology North Bangkok (KMUTNB), in 2011, and the Ph.D. degree from TU Darmstadt, in 2016. His Ph.D. dissertation is titled ‘‘Advanced Monitoring in P2P Botnets.’’ He has been a Senior Lecturer with the National Advanced IPv6 Centre (NAv6), Universiti Sains Malaysia, since 2016. He has also been a Senior Researcher/Postdoctoral Researcher with the Telecooperation Group, TU Darmstadt, since July 2019. He is currently working on several cyber-security projects and groups, such as the National Research Center for Applied Cybersecurity (ATHENE), formerly known as the Center for Research in Security and Privacy (CRISP).

References

Abri

Jamouss

Kraiem

, et al. (2020) Perlcol: a framework for personalized e-learning with social collaboration support. International Journal of Computing and Digital Systems 9(3): 483–493. DOI: 10.12785/IJCDS/090312

Aeiad

Mezaine

(2019) An adaptable and personalized E-learning system applied to computer science Programmes design. Education and Information Technologies 24: 1485–1509.

Aguilar

Buendia

Pinto

, et al. (2022) Social learning analytics for determining learning styles in a smart classroom. Interactive Learning Environments 30(2): 245–261. DOI: 10.1080/10494820.2019.1651745

Ashraf

Manickam

Karuppayah

(2021) A comprehensive review of course recommender systems in e-learning. Journal of Educators Online 18(1): 23–35.

Ashraf

Selvakumar

Shankar

, et al. (2023) An intelligent E-learning course recommendation framework based on student learning style. Journal of Educators Online 20(1): 33–42. DOI: 10.9743/jeo.2023.20.1.1

Bachiri

Mouncif

(2023) Artificial intelligence system in aid of pedagogical engineering for knowledge assessment on MOOC platforms: open EdX and moodle. International Journal of Emerging Technologies in Learning 18(5): 144–160. DOI: 10.3991/ijet.v18i05.36589

Baidada

Mansouri

Poirier

(2019) Personalized e-learning recommender system to adjust learners’ level. EdMedia+ innovate learning. Association for the Advancement of Computing in Education (AACE)., 1353–1357.

Bhaskaran

Santhi

(2019) An efficient personalized trust based hybrid recommendation ( TBHR ) strategy for e-learning system in cloud computing. Cluster Computing 22: 1137–1149. DOI: 10.1007/s10586-017-1160-5

Chandra

Verma

Lim

, et al. (2022) Personalization in personalized marketing: trends and ways forward. Psychology and Marketing 39(8): 1529–1562. DOI: 10.1002/mar.21670

10.

Chen

Yin

, et al. (2020) Enhanced learning resource recommendation based on online learning style model. Tsinghua Science and Technology 25(3): 348–356. DOI: 10.26599/TST.2019.9010014

11.

Dhananjaya

Anjanabhargavi

Vijayalaxmi

, et al. (2024) A digital recommendation system for personalized learning to enhance online education: a review. IEEE Access 1. DOI: 10.1109/ACCESS.2024.3369901

12.

El-Bishouty

Aldraiweesh

Alturki

, et al. (2019) Use of felder and silverman learning style model for online course design. Educational Technology Research & Development 67(1): 161–177. DOI: 10.1007/s11423-018-9634-6

13.

Khanal

Prasad

Alsadoon

, et al. (2020) A systematic review: machine learning based recommendation systems for e-learning. Education and Information Technologies 25(4): 2635–2664. DOI: 10.1007/s10639-019-10063-9

14.

Khanse

Bhandari

Singru

, et al. (2020) Comparative study of genetic algorithm and artificial neural network for multi-class classification based on type-2 diabetes treatment recommendation model, In 6th International Conference on Parallel, Distributed and Grid Computing (PDGC), Himachal Pradesh, India, 6-8 Nov 2020, pp. 538–543. DOI: 10.1109/PDGC50313.2020.9315837

15.

Kulkarni

Rai

Kale

(2020) Recommender system in eLearning: a survey. In: Proceeding of International Conference on Computational Science and Applications. Berlin, Germany: Springer nature, 119–126. DOI: 10.1007/978-981-15-0790-8_13

16.

Laksitowening

Yanuarifiani

Wibowo

YFA

(2016) Enhancing e-learning system to support learning style based personalization. Proceeding - 2016 2nd International Conference on Science in Information Technology, ICSITech 2016: Information Science for Green. 2nd International Conference on Science in Information Technology (ICSITech): IEEE, 329–333.

17.

Mahmood

Ismail

(2021) The effect of hull fineness ratio and fin parameters on the optimization of tethered aerostat. Aircraft Engineering & Aerospace Technology 93(4): 690–700. DOI: 10.1108/AEAT-04-2020-0071

18.

Martin

Bolliger

Flowers

(2021) Design matters: development and validation of the online course design elements (OCDE) instrument. International Review of Research in Open and Distance Learning 22(2): 46–71. DOI: 10.19173/irrodl.v22i2.5187

19.

Monsalve-Pulido

Montoya

Salazar

(2024) Autonomous recommender system architecture for virtual learning environments. Applied Computing and Informatics 20(1/2): 69–88. DOI: 10.1016/j.aci.2020.03.001

20.

Nafea

Siewe

(2019) On recommendation of learning objects using felder-silverman learning style model. IEEE Access 7: 163034–163048. DOI: 10.1109/ACCESS.2019.2935417

21.

Pavlidis

(2019) Recommender systems, cultural heritage applications, and the way forward. Journal of Cultural Heritage 35: 183–196. DOI: 10.1016/j.culher.2018.06.003

22.

Pellegrini

Masquelier

(2021) Fast threshold optimization for multi-label audio tagging using surrogate gradient learning, In IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings (ICASSP), Toronto, ON, Canada, 6-11 June 2021, pp. 651–655. DOI: 10.1109/ICASSP39728.2021.9414091

23.

Study.com (2022) Available at: https://study.com/academy/lesson/teaching-methods-typesstrategies.html

24.

Syed

Nair

SSK

(2018) Personalized recommendation system for advanced learning management systems. ACM International Conference Proceeding Series: 90–95. DOI: 10.1145/3268891.3268899

25.

Thongchotchat

Kudo

Okada

, et al. (2023) Educational recommendation system utilizing learning styles: a systematic literature review. IEEE Access 11: 8988–8999. DOI: 10.1109/ACCESS.2023.3238417

26.

Varniab

Hung

Sharghi

(2019) Classification of multiclass datasets using genetic programming, In Proceedings of the 2019 Research in Adaptive and Convergent Systems (RACS), Chongqing China, September 24-27, 2019, pp. 76–82. DOI: 10.1145/3338840.3355656

27.

Yadav

Sohal

(2017) Comparative study of different selection techniques in genetic algorithm. International Journal of Energy Sector Management 6(3): 174–180. Available at: https://www.researchgate.net/publication/325011628