Consequential effects of using competing perspectives to predict learning style in e-learning systems

Abstract

The learning processes have been significantly impacted by technology. Numerous learners have adopted technology-based learning systems as the preferred form of learning. It is then necessary to identify the learning styles of learners to deliver appropriate resources, engage them, increase their motivation, and enhance their satisfaction and learning outcomes. Adopting mixed method, this study evaluated the effects of using collaborative and automatic perspectives to predict VARK learning styles. Participants were first-year undergraduate Computer Science students. Questionnaires were administered and activity/event data were collected for analysis. Using Naive Bayes, J48, OneR, and SMO classifiers to extract patterns in activity/event data, the experimental results of the automatic perspective indicated improved classification of learning styles and performance of students compared with the collaborative technique. The prediction relevance of multimodal and read/write learning styles and accuracy of classification were better than other learning styles. Kinesthetic was not automatically identified. The students have different learning preferences in the same discipline; however, multimodal and read/write learning styles were dominant in both perspectives. Students who understood varied and simple formats of learning imagery objects were more likely to respond effectively to related assessments, and those who received varieties of course contents and information had their performance improved.

Keywords:

• Clustering and classification algorithms
• collaborative and automated techniques
• e-learning
• predicting learning style
• VARK learning styles

1. Introduction

Learning style is an educational setting that is conducive for a learner to learn. It is the preferred manner in which a learner sees, processes, comprehends, and remembers knowledge. It is an intrinsic trait that is impacted by the environment, experiences, and changes. Learning styles provide clues about how learners see, engage with, and respond to their learning situations. It is best to periodically assess the learner’s preferred method of learning using a specific learning style model. The VARK (Visual, Auditory, Read/Write, Kinesthetic) model proposed by N. D. Fleming (1995) is often used to identify learning styles of learners. Visual learners learn best when they see things, prefer graphic illustration and observation, and use imagery and diverse formats, fonts and colours to highlight main ideas. Auditory learners acquire knowledge and become skilled at something by listening, Read/write learners learn well through repetition of written words and retain information well by having notes, and Kinesthetic learners learn through moving, touching, and doing by eye-hand collaboration. Clustering and classification algorithms are frequently used to automatically detect learning styles of learners on e-learning systems. Among these are Naive Bayes, J48, One Rule (OneR), and Sequential Minimal Optimization (SMO) classifiers. Clustering algorithms compare membership using iterative strategy to find the distance value and compare with the threshold value. Also, variety of techniques such as automated and collaborative are used to facilitate the identification of learning styles. Automated technique is considered more reliable because the actual behaviour of a learner is used as input to identify learning styles.

Moodle-based e-learning system was chosen because it is an open-source framework that permits enhancement, most popular learning management system (LMS) among educational organizations, and supports provision of e-learning solutions with limited infrastructure and resources. Context-relevant contents such as learner model, group model, domain model, and adaptation model were essential for categorizing the learning styles. Sharable Content Object Reference Model packages were used to achieve interoperability and displaying course content. Internal attributes of the e-learning website including checking errors, browser compatibility, and loading time were assessed using HTML Toolbox and Webpage Analyzer. Data and other related usage logs were captured, serialized, mapped automatically to each category of VARK learning styles and stored on WAMP server (MySQL database). Adaptation content stored include learning object, navigation/link support, structure, and presentation. HTML/CSS, JavaScript and PHP were used to develop and improve dynamic and aesthetic features of the Moodle. Personalized user interfaces were generated for each identified learning style category and course contents recommended. WEKA software by Waikato University in New Zealand was used to analyse, cluster and classify activity/event logs into learning styles.

The learning and teaching processes have been significantly impacted by technology. Numerous students have adopted technologically based learning systems as the preferred form of learning (Obeng et al., 2020). This has made it critical for educational institutions to identify the learning styles of learners to deliver appropriate resources, engage students, increase their motivation, and enhance their satisfaction and learning productivity in e-learning systems (Kangas et al., 2017). Identifying the right learning styles of students and delivering appropriate learning resources could help improve learning outcome of students (Raj & Renumol, 2022). Easy understanding of textual learning objects and easy to access and understand varieties of learning objects can also contribute to the improvement in learning outcome of the students (Bali et al., 2022). Students who understand varied and simple formats of learning imagery objects are more likely to respond effectively to related assessments, and those who receive varieties of course contents and information can have their learning performance improved (Bali & Rozhana, 2022).

The main objective of this study was to evaluate the effects of learning style prediction on learning outcome using competing perspectives. Collaborative perspective embraced questionnaires to identify learning styles while clustering and classification algorithms were used to automatically detect learning styles of students. We pre-processed student learning behaviour after extracting it from learner logs. This processed data served as input to machine learning classifiers to identify the learner’s preferred method of learning. In this paper, section 2 discusses the study’s background while section 3 focuses on related work. The method is described in Section 4. While section 4 discusses and corroborate results, section 5 presents conclusion. Finally, research contribution, limitations and future work are discussed at section 6.

2. Background

2.1. Learning style

The automated identification of learners’ learning preferences has attracted research attention as a way to provide a learner-centred learning environment. Each learner has a preferred method of learning based on personality and contextual conditions. For instance, some learners may grasp experiment instructions by listening to them, while others need to physically do the experiment. The educational systems should take into account these variations in learner s’ learning styles to improve the learning process. The phrase “learning style” has received several meanings. Learning style is considered an educational setting that is conducive for a learner to learn. Learning style is described as the distinctive preferences and strengths in how information is ingested and processed by learners (Felder, 1996). The preferred manner in which a learner sees, processes, comprehends, and remembers knowledge is referred to as a learning style. Thus, related attitudes and behaviours that facilitate and improve learning irrespective of the circumstance is considered as learning style.

The main focus of learning styles is how people choose to learn, not what they learn. When someone attempts to learn something new, their preferences might change. When it comes to learning, each person has unique preferences; this is an intrinsic trait that is impacted by the environment, experiences, and changes. According to Law et al. (2010), factors that focus on environmental setting including “clear direction”, “reward and recognition”, “punishment”, and “social pressure and competition” are extrinsic factors that motivate learners to partake in e-learning. Learning-teaching process, online course environment/technical infrastructure, and measurement and evaluation (Selvi, 2010) are considered other extrinsic motivating factors. These factors result in flexibility, deliverability, freedom and independency of use of online environments (Şahin et al., 2022). Extrinsic motivation features of e-learning systems, such as usefulness and convenience of use, can motivate students more effectively (Şahin et al., 2022). Numerous models of learning style have been proposed, however, eminent educational theorists agree that every person has a preferred method of learning that the learning environment must accommodate. These learning styles provide unmistakable clues about how learners see, engage with, and respond to their learning situations. Learning preferences are not dynamic; they do not alter often. However, there is always a better method to learn various topics in various ways. It is best to periodically assess the learner’s preferred method of learning using a specific learning style model.

2.2. VARK learning style

Depending on the concept used for analysis and study, different learning styles can be identified in learners. The VARK (Visual, Auditory, Read/Write, Kinesthetic) model proposed by N. D. Fleming (1995) is used to classify learning style of learners. The VARK model is made up of 16 items, presented as questionnaire. Visual learners learn best when they see things, prefer graphic illustration and observation (N. Fleming, 2011) and uses imagery and diverse formats, fonts and colours to highlight main ideas. Auditory learners acquire knowledge and become skilled at something by listening while read/write learners learn well through repetition of written words and retain information well by having notes. Kinesthetic learners learn through moving, touching and doing by eye-hand collaboration. The bimodal, trimodal and multimodal use combinations of either four categories.

2.3. Perspectives of identifying learning style

To determine the different learning styles of students, researchers have used a range of approaches and based their research on a number of learning theories. The automated and collaborative methods are the two major methods used to detect learning styles. To determine learning preferences for the collaborative approach, a questionnaire is employed. Collaborative technique is simple to use (Kolekar et al., 2017) and produces accurate learning style identification results than the automatic approach (Pérez-Marín et al., 2022). Nevertheless, the collaborative approach has drawbacks since certain patterns that are behavioural indicators of learning styles may not be captured when respondents are filling out the survey, which ultimately might undermine the validity of the findings (Pérez-Marín et al., 2022). The collaborative technique relies only on basic rules to determine learning style.

Using the automated method to determine learning styles is thus necessary since it is thought to be able to assess learning and provide prompt responses depending on the requirements of the learner. Result from automated technique is considered more reliable because the actual behaviour of a learner is used as input (Hasibuan et al., 2019). However, much effort is needed to obtain the behavioural patterns of learners. Different categorization methods, including Bayesian networks, fuzzy clustering, decision trees, and hidden Markov have been applied by academics employing data-driven approach to identify learning styles automatically. All of these techniques developed classifiers utilizing the retrieved data after extracting certain variables from learner behaviour (Bernard et al., 2017).

This study used VARK model proposed by N. D. Fleming (1995) to determine learning styles of students. Using VARK model is an effective way to determine learning styles (Pérez-Marín et al. (2022); El-Sabagh (2021); and Chrysafiadi et al. (2019)). Stojanova et al. (2017) and Díaz et al. (2018) used VARK model to identify learning styles of students who were pursuing Computer Science courses including Data Structure and Algorithms. They indicated benefits of using the VARK model. Using VARK model, this study complemented the collaborative and automated methods to detect learning styles of students.

2.4. Clustering and classification algorithms

This section presents how the Expectation—Maximization clustering and Naive Bayes, J48, OneR, and SMO classification algorithms are operationalized.

2.4.1. The Expectation–Maximization (EM) clustering algorithm

The EM clustering algorithm is chosen since it produces maximum-likelihood (ML) estimates of parameters when there is a many-to-one mapping (Moon, 1996) similar to the underlying observation of this study. Finding the unknown underlying variables using the current estimates of the parameters to determine the probability of cluster element(s) $\mathrm{P}\left(C_j|x_k\right)$ is achieved through the expectation step of EM. Where $x_k$ represents attribute vector composition of an element, the relevance degree of the points of each cluster is given by the likelihood of each element attribute in comparison with the attributes of the other elements of cluster $C_j$. Given below:

(1) $\mathrm{P}\left({\mathrm{C}}_{\mathrm{j}}|\mathrm{x}\right)=\frac{|{\sum }_{\mathrm{j}}(\mathrm{t}){|}^{-\frac{1}{2}}\mathrm{e}\mathrm{x}{\mathrm{p}}^{\mathrm{n}\mathrm{j}}{\mathrm{P}}_{\mathrm{j}}(\mathrm{t})}{\sum _{\mathrm{k}=1}^{\mathrm{M}}|{\sum }_{\mathrm{j}}(\mathrm{t}){|}^{-\frac{1}{2}}\mathrm{e}\mathrm{x}{\mathrm{p}}^{\mathrm{n}\mathrm{j}}{\mathrm{P}}_{\mathrm{j}}(\mathrm{t})}$(1)

x represents input dataset, M for the total number of clusters, and t an instance with initial instance value of zero.

The maximization step estimates the new parameters of the probability distribution of each class by computing the mean ${\mu }_j$ of class j obtained through the mean of all points in function of the relevance degree of each point. The mean of probabilities $C_j$ computes the probability of occurrence of each class considering the relevance degree of each point from the class. Given below:

(2) $P_j\left(t+1\right)=\frac{1}{N}{\mathrm{\Sigma }}_{k=1}^M\mathrm{P}\left(C_j|x_k\right)$(2)

x represents input dataset, M for the total number of clusters, and t an instance with initial instance value of zero.

Bayes theorem is used to calculate the covariance matrix at each iteration where EM algorithm increases the likelihood function until a point of maximum likelihood (local) is reached (Moon, 1996). The significance of Bayes’ Theorem is relating the “direct” probability of P_E(H) to the “inverse” probability of P_H(E) of the conditional data that results in:

${\mathrm{P}}_E\left(H\right)=\left[\mathrm{P}\left(H\right)/\mathrm{P}\left(E\right)\right]{\mathrm{P}}_H\left(E\right)$

where

P_E is the probability function, E is the logical consequence (P_E(H) = 1 if E entails H), preservation of certainties (If P(H) = 1, then P_E(H) = 1), Mixing (P(H) = P(E)P_E(H) + P(~E)P_~E(H)).

The EM Algorithm 1 below shows step-by-step procedure to automatically group together learning styles of students using the feature values of VARK and learning sequences. Thus, members are assigned specific sequence values that maps to a singular cluster centre. Members of a group further move toward the exact cluster centre when the sequence values are closest. Cluster centres are revised after comparing membership using iterative strategy to find the distance value and compare with the threshold value. The sequence of a learner is grouped into a cluster where the distance value is less than the threshold value. Though a sequence can belong to more than one cluster, summing all values of members’ sequence should be equal to one. At the end, all sequences are categorized into four clusters of learning styles and assigned label as per VARK. Where sequence of a learner belongs to more than one cluster, it is classified as multimodal.

Algorithm 1: Using EM algorithm to converge parameter estimate

Step 1: Select anInitial parameter, ${\theta }^{\left[0\right]}$ set s = 0       

Step 2: In Expectationstep, estimate unobserved data using ${\theta }^{\left[s\right]}$

Step 3: In Maximum step, compute maximum likelihood estimate of parameter ${\theta }^{\left[s+1\right]}$ using estimated data

Step 4: If s = s +1 is converged (entire data elements are clustered), then STOP, else go to step 2

2.4.2. Classification algorithm

All sequences labelled using the EM algorithm and new sequence of each learner are used as input to train the classification algorithms of Naive Bayes, J48, OneR, and SMO. It is necessary to classify new sequence of learner’s activities according to the four categories of VARK learning styles. In automatic detection of learning styles, classification algorithms are commonly used (Feldman et al., 2014) to learn from previous examples in order to determine the nearest target attribute.

For example, J48 inductive learning is defined:

K={$k_{1,}{k}_{2,\dots ,}{k}_n$}: a set of training examples

S={$s_{1,}{s}_{2,\dots ,}{s}_m$}: a set of descriptive attributes

p: the target attribute

Each training example $k_{1,}$ is represented by a vector <$x_{1,}{x}_{2,\dots ,}{x}_m{a}_i$>, where $x_{1,}{x}_{2,\dots ,}{x}_n$ denotes a legal value of attribute $s_{1,}{s}_{2,\dots ,}{s}_{m,}$ and $a_i$ is a legal value of the target attribute p.

2.4.3. Validation of algorithm

Classification of students’ learning styles often apply Naive Bayes, J48, OneR, and SMO (Şanlı et al., 2020). The study adopted Naive Bayes, J48, OneR, and SMO classifiers that use gain ratio data reduction/filtering method to select nodes and test relevant attributes. Rules of classification produced by these classifiers are easy to understand for pattern recognition and classification (Nakra & Duha, 2019), simple to apply, and works efficiently where many attributes are included in classification (Feldman et al., 2014).

Gain ratio normalizes information gain using the formula:

SplitInfo_A(S) $=-{\sum }_{i=1}^v\left(\left|s_i|/|s\right|\right)\mathrm{l}\mathrm{o}{\mathrm{g}}_2$ $\left(\left|s_i|/|s\right|\right)$

where S is the training data set split into v partitions that corresponds to v outcomes of a test on the attribute A. Selecting splitting attribute is based on attribute with highest gain ratio (Han & Kamber, 2001), hence, gain ratio formula is given as:

Gain Ratio(A) = Gain (A)/SplitInfo_A(S)

3. Related work

Throughout the past two decades, learner behaviour modelling has attracted a lot of interest (Abyaa et al., 2019). Learners from different fields have different ways of learning and require educational materials that suit their learning styles (Churngchow et al., 2020). Khan et al., (2018) conducted a study that included the typical behaviours, learning preferences, recall and retention rates, efficiency, and performance of students using different learning styles. These authors reported successful identification of learning styles. Researchers draw conclusion on e-Learning systems that use online models as very simple to use, are more motivating for participating learners, and assist learners in improving their academic performance (Tawafak et al., 2019). Lwande et al. (2021) recommend automatic matching of online courses contents to students’ individual learning preferences and cognitive characteristics.

Researchers have proposed and developed many algorithms to automatically predict learning styles of learners on e-learning systems. Using Kolb’s learning style theory, Kalhoro et al. (2016) proposed a method to use Data Mining and a Decision Tree (J48) classifier to automatically identify students’ learning preferences from their blogs. Bayesian network was utilized by Balasubramanian and Anouncia (2018) to determine the learning preferences of students. They found a number of behaviours that might be useful in determining learning styles. After training the Bayesian network, the likelihood that each student prefers a certain learning style was calculated. In a study, Ferreira et al. (2019) defined students’ learning profiles using the Felder-Silverman Learning Style Model by using machine learning techniques to analyse data from Moodle. The literature-based and data-driven approaches was used by Karagiannis and Satratzemi (2019), Prior Knowledge and Rule-based by Sweta and Lal (2017), and data-driven and Artificial neural networks by Gomede et al. (2020) to identify learning styles. Merlini and Rossini (2021) used WEKA to implement the SMO algorithm to train a support vector classifier. Jen and Lin (2021) adopted OneR algorithm to determine the accuracy of classification algorithms for data prediction in machine learning applications. Ouatik et al. (2022) used C4.5/J48; Hasibuan et al. (2019) used C4.5 and Naïve Bayes; and Jagadeesan and Subbiah (2020) used Naïve Bayes as automated methods to detect the learning styles of students.

According to Pérez-Marín et al. (2022); El-Sabagh (2021); and Chrysafiadi et al. (2019), using VARK questionnaire is effective collaborative approach to determine learning styles. Stojanova et al. (2017) and Díaz et al. (2018) successfully used VARK model to identify learning styles of students who were pursuing Computer Science courses. Using artificial neural network (data-driven approach) and VARK (collaborative approach), Chrysafiadi et al. (2019) successfully presented personalized and adaptive content to learners. Pantho (2016) presented a method for categorizing learners’ VARK learning preferences using J48 algorithm. A questionnaire that received responses from 1,205 students was used to gather information on learners’ learning methods. J48 was then used to classify the obtained data. In order to assure effective resource recommendations, neural network was employed in the automatic identification of the learning styles.

4. Methods

The study adopted a mixed method (collaborative and automated) with the intent of comparing learning styles and performance levels of students. Survey was used to obtain data from a segment of students that share similar characteristics (Creswell, 2012). Using convenience sampling approach, only first-year Computer Science students studying Ethical and Legal Implications of Computing at a tertiary educational institute in Ghana that were accessible and available (Gravetter & Forzano, 2012) were selected. This helped in minimising large variations in cognitive ability, age and culture among the participants. Students were tasked to use e-learning system (Moodle) between February and March, 2022. The intents were that: all students should have knowledge of using Moodle; the authors should be able to assess the level of memorability of students; the authors would be able to understand the usability and efficiency levels of the students; and authors would be able to administer questionnaire to acquire learning styles. The active default learning style was assigned to all students when log in for the first time.

Using VARK version 8.01, a questionnaire consisting of 16 questions with four (4) possible answers each on VARK learning styles was developed and administered through Moodle platform between April 2022 and June 2022 after fine-tuning with experts’ views. Out of 214 students in class, 190 (response rate of 86%) completed the online survey. Employing descriptive statistics, frequencies of occurrences were obtained to capture, understand, and summarize the responses of students and their learning styles. Within this same period, the behaviour/usage (activity/event logs) data of the 190 students were automatically (data-driven approach) captured, analysed, clustered and classified into various learning styles using WEKA. This was done to corroborate results from collaborative method.

4.1. Context-relevant contents

Context-relevant contents including learner model, group model, domain model and adaptation model are considered critical success factors when categorizing learning styles in e-learning system (Tmimi et al., 2019). Learner model includes learners’ backgrounds, knowledge, goals/tasks, previous learning experience, preferences, interests, and interaction style (Zamecnik et al., 2022). Learner data stored were learner profile (name, age, email, username and password, and educational background); learner preference (learning style, language, interest, content preference, interaction style, goal/task, environment/user platform); and learner portfolio (academic performance—Student ID, content studied, date and time, state of completion, chronometric (accumulated usage time), and knowledge (score, grade)). For group model, categories of VARK learning styles were stored as group data. Only registered users were able to access the learning resources on Moodle. Users had access to all resources when log into the system. Learning components including announcement, examination, assignments, quiz, exercise, references, topic list, discussion forum, email, topic level (basic, intermediate, advanced) were defined for the e-learning portal alongside learning contents (text, pptx, pdf, videos, images, charts, graphs, diagrams, flowcharts, shapes, voice, links, colours, etc.); contents indexes, links of topic, learner profile; search/filter and assignments submission/download options.

4.2. Automatic categorization of learning styles of students

Using Moodle framework, an e-learning application that has the capability of capturing the usage data of students was developed. Following the data-driven approach and using clustering and classification techniques, activity logs showing learning components and learning objects (average viewing frequency of learning object, average viewing time (minutes) of learning object, web page id, student id) accessed by students for a stipulated period of two months were serialized and mapped automatically to each category of VARK learning styles. Tagging results with personal profiles, mainly the user identifier (in this case, student id) resulted in automatic update of results in the database. The Expectation—Maximization clustering and Naïve Bayes, J48, SMO, and OneR classification algorithms used to categorize the learning styles of students automatically are presented in the subsequent sections. Table 1 shows the main attributes, their descriptions and possible values of the dataset used.

Table 1. Attributes, description and possible values

Attributes	Description	Possible Values
stuID	Student identifier	200–397
platType	Operating system used	Windows (1), Linux (2), MacOS (3), Android (4)
pageID	Unique ID for each web page	101, 102, 103, 104, 105
learningStyle	Learning style of users	Video (1), Audio (2), ReadWrite (3), Kinesthetic (4), Multimodal (5)
assGrade	Average grade from quizzes, assignments, examinations	Excellent (≥80), Credit (70–79), Good (60–69), Pass (50–59), Fail (0–49)
avgAssTime	Average time (minutes) spent on quizzes, assignments, examinations	0–25 (1), 26–40 (2), 41–60 (3)
AvgLearningObjectAccessFreq	Average frequency of accessing learning objects	Once (1), twice (2), three times (3)
AvgLearningObjectAccessTime	Average time spent on accessing learning objects	1–10 (1), 11–20 (2), 21–30 (3), 31–40 (4), over 40 minutes (5)
avgTextView	Average frequency of accessing text	Once (1), twice (2), three times (3)
avgVideoView	Average frequency of accessing video	Once (1), twice (2), three times (3)
avgAudioListen	Average frequency of accessing audio	Once (1), twice (2), three times (3)
avgPracticalDone	Average frequency of accessing practical	Once (1), twice (2), three times (3)
avgTexTime	Average time spent on accessing text	1–10 (1), 11–20 (2), 21–30 (3), 31–40 (4), over 40 minutes (5)
avgVideoTime	Average time spent on accessing video	1–10 (1), 11–20 (2), 21–30 (3), 31–40 (4), over 40 minutes (5)
avgAudioTime	Average time spent on listening to audio	1–10 (1), 11–20 (2), 21–30 (3), 31–40 (4), over 40 minutes (5)
avgPracticalTime	Average time spent on doing practical	1–10 (1), 11–20 (2), 21–30 (3), 31–40 (4), over 40 minutes (5)

5. Experimentation result and discussion

This study focused on evaluating the effects of learning style prediction on learning outcome using competing perspectives. Choosing Moodle-based e-learning system was appropriate because it is an open-source framework that permits enhancement, most popular learning management system (LMS) among educational organizations, supports provision of e-learning solutions with limited infrastructure and resources, and has over 359.6 million users in 242 countries (Moodle, 2023). Additionally, Moodle can be deployed as Mobile Application, Self-Hosted Cloud-based, or Self-Hosted System; on varied platforms (Unix, Linux, Mac, Windows), and runs on several browsers.

The following data and other related usage logs were captured and stored on WAMP server, specifically into MySQL database. Domain model focused on the level of difficulty of the content (such as easy, difficult), time required to complete learning object, and prior knowledge to be able to perform the learning object (prerequisite). Stored domain level data were the course, structure of links, nodes navigation content; error checking, browser compatibility, and page loading time. Adaptation content stored include learning object consisting text and multimedia – (graphic: image, charts, symbol), (video: audio, animation), (text: word processor, PowerPoint, pdf, etc), and XML (Web, Sharable Content Object Reference Model-SCORM); navigation/link support (link hiding, sorting, annotation, direct guidance, new links, similarity-based links, adaptive links); structure (how the same page is displayed in different structures and forms); and presentation (same content in customized styles—text, colour, font, etc.).

The framework that Moodle provides in building websites is static with limited functionalities. HTML/CSS, JavaScript and PHP were used to develop and improved dynamic and aesthetic features of the Moodle. Internal attributes of the e-learning website including checking errors, browser compatibility, and loading time were assessed using HTML Toolbox and Webpage Analyzer. On the portal, personalized user interfaces were generated for each identified learning style category and course contents were recommended. Sharable Content Object Reference Model packages were used to achieve interoperability and displayed course content. Designing interfaces of web pages factored in Nielsen’s 10 heuristic principles for interface design (Nielsen, 2020), usability attributes (Nielsen, 2012), and Web Content Accessibility Guidelines (WCAG) 2.1 (W3C, 2018).

Results of data-driven technique obtained from WEKA was compared and corroborated with the results of the collaborative technique. WEKA software by Waikato University in New Zealand was utilized to analyse, cluster and classify activity/event logs (usage data) into learning styles. Operation of WEKA is centred on machine learning and data mining. WEKA has several functions that facilitate classification, performing of regression analysis, clustering, application of association rules, selection, and visualization of data.

5.1. Results of the collaborative technique

Figure 1 shows the VARK learning styles of students obtained through questionnaire consisted of 16 questions with four possible answers each. All VARK learning styles and multimodal that uses combinations of either of the four categories were identified from the questionnaire.

Figure 1. VARK learning styles of students.

A column bar graph plotting the frequencies of five learning styles of students obtained through questionnaire. Multimodal is the dominant learning style.

Clearly, multimodal (46.7%) was the dominant learning style among the students while visual learning style (9.2%) was the least.

Prior to using the activities logs to automatically identify learning style, students were tasked to perform learning and assessment activities on the Moodle for a period of 2 months. Figure 2 represents results of students that were obtained from the assessments conducted within this 2-month period. The failure rate was high in the assessment, 56 students representing 30.4%.

Figure 2. Assessment results of students prior to automatic identification of learning styles.

A line graph plotting frequencies and assessment scores of students prior to automatic identification of learning styles. Grading ranges from fail to excellent, with students obtaining more failure grades.

5.2. Results of the automated technique

This section presents the descriptive and WEKA results of the automated technique.

5.2.1. Descriptive statistics of collaborative and automated techniques

Figure 3 shows frequencies of learning styles of the collaborative and automated techniques for easy comparison.

Figure 3. Comparing learning styles frequencies of the collaborative and automated techniques.

A column bar graph to compare frequencies of five learning styles of students obtained from collaborative and automated techniques. Kinesthetic learning style has zero frequency from automated techniques.

Automatic detection of visual, auditory, and read/write were higher than those identified from the questionnaire. Visual increased by 50%, auditory 49%, read/write 60% and multimodal reduced by 34%. Kinesthetic was not automatically identified.

Figure 4 represents results of assessment of students post automatic identification of learning styles. The performance of students significantly improved. Students who scored 80 and above increased from 42 (see Figure 2) to 63, while the number of students that scored least reduced to 33 from 56 (see Figure 2).

Figure 4. Assessment results of students post automatic identification of learning styles.

A line graph plotting frequencies and assessment scores of students after automatic identification of learning styles. Grading ranges from fail to excellent, with students obtaining more excellent grades.

The experimental results indicated improved classification of learning styles and performance of students which are similar to the findings of Kolekar (20187). This improvement could have resulted from identifying the right learning styles of students and delivering to them appropriate learning resources. This affirms the position of Mirza and Khurshid (2020) and Kim et al. (2018) where identified right learning styles facilitated adapting learning resources to the needs of the learners that eventually improved their learning processes and performance. Similar to findings of Mirza and Khurshid (2020), the students had different learning preferences in the same discipline.

Although students pursuing computer-related programmes are distributed across visual, auditory, read/write, and kinesthetic learning styles (Mirza & Khurshid, 2020), kinesthetic was not automatically detected in this study. This was so because the ethical and legal implications of computing which was used in the identification of learning styles is a reading course, hence, multimodal and read/write learning styles were dominant (Mirza & Khurshid, 2020).

In line with the findings of Cidral et al. (2018), easy understanding of textual learning objects significantly contributed to the improvement in learning outcome of the students. Understanding varied and simple formats of learning imagery objects and easy to access and understand varieties of learning objects (Obeng et al., 2020) also contributed to the improvement in learning outcome of the students. These confirm research findings of Abdullah and Rowley (2018) and Cidral et al. (2018). Students who understand varied and simple formats of learning imagery objects are more likely to respond effectively to related assessments. Varieties of course contents and information often fit the needs of a learner (Uppal et al., 2017). Students who received varieties of course contents and information had their performance improved. Flexible to learn visual and textual learning objects and availability of different learning objects (Kim et al., 2018) are more likely to learn well, hence, increasing in the number of multimodal learning styles.

5.2.2. WEKA results

On the Moodle, personalized user interfaces were generated for each identified learning style category and course contents recommended. The logs of activities performed by students for a period of 2 months were analysed using WEKA software. We run Naive Bayes, J48, OneR, and SMO classifiers with 10-fold cross-validation each. This was done to help validate the data, and to compare and affirm results. Table 2 illustrates the summary of performance evaluation of the classifiers, while Table 3 shows accuracy of evaluation metrics of the classifiers per all classified learning styles.

Table 2. Summary of performance evaluation of the classifiers

	NaiveBayes	J48	OneR	SMO
Correctly classified instances	133 (67.1717 %)	152 (76.7677%)	157 (79.2929%)	137 (69.1919%)
Incorrectly classified instances	65 (32.8283 %)	46 (32.8283%)	41 (20.7071%)	61 (30.8081%)
Kappa statistics	0.5711	0.6962	0.7284	0.5924
Mean absolute error	0.1624	0.1325	0.0828	0.2653
Root mean squared error	0.3096	0.2782	0.2878	0.3524
Relative absolute error	42.527 %	42.8545 %	26.7939 %	55.806 %
Root relative squared error	48.7639 %	50.7887 %	53.2288 %	59.6575 %
Total number of instances	198	198	198	198

Table 3. Accuracy of evaluation metrics of the classifiers per classified learning styles

Classifier	Learning style	Evaluation metrics
		TP Rate	FP Rate	Precision	Recall	F-Measure	MCC	ROC	PRC
NaiveBayes	ReadWrite	0.765	0.054	0.830	0.765	0.796	0.730	0.925	0.823
	MultiModal	0.879	0.064	0.850	0.879	0.864	0.807	0.915	0.835
	Kinesthetic	0.000	0.017	0.000	0.000	0.000	−0.038	0.474	0.087
	Visual	0.558	0.174	0.471	0.558	0.511	0.362	0.792	0.521
	Auditory	0.655	0.107	0.514	0.655	0.576	0.498	0.915	0.611
J48	ReadWrite	0.961	0.082	0.803	0.961	0.875	0.833	0.928	0.726
	MultiModal	0.879	0.071	0.836	0.879	0.857	0.796	0.877	0.706
	Kinesthetic	0.059	0.039	0.125	0.059	0.080	0.029	0.463	0.092
	Visual	0.581	0.052	0.758	0.581	0.658	0.586	0.801	0.710
	Auditory	0.897	0.053	0.743	0.897	0.813	0.782	0.945	0.694
OneR	ReadWrite	0.961	0.082	0.803	0.961	0.875	0.833	0.940	0.782
	MultiModal	0.879	0.064	0.850	0.879	0.864	0.807	0.908	0.783
	Kinesthetic	0.000	0.000	?	0.000	?	?	0.500	0.086
	Visual	0.651	0.045	0.800	0.651	0.718	0.655	0.803	0.597
	Auditory	1.000	0.077	0.690	1.000	0.817	0.798	0.962	0.690
SMO	ReadWrite	0.922	0.129	0.712	0.922	0.803	0.735	0.915	0.708
	MultiModal	0.879	0.064	0.850	0.879	0.864	0.807	0.909	0.795
	Kinesthetic	0.000	0.017	0.000	0.000	0.000	−0.038	0.585	0.112
	Visual	0.651	0.129	0.583	0.651	0.615	0.502	0.825	0.512
	Auditory	0.379	0.059	0.524	0.379	0.440	0.368	0.772	0.342

All the classifiers (Naive Bayes, J48, OneR, and SMO) have good agreement for Kappa with values ranging from 0.5711 to 0.7284, see Table 2. Each model correctly classified the training data in 0 second. The performances of all classifiers are better than baseline (Root relative squared error) accuracy values as shown on Table 2. However, OneR and J48 classifiers performed better and correctly classified instances than SMO and Naïve Bayes. The models are better with Mean Absolute Error values ranging between 0.0828 to 0.2653. The Precision (ranging from 0.712–0.850) that represents a true positive rate of prediction and Recall (ranging from 0.712–0.961), a proportion of retrieved relevant instances of all classifiers indicate better prediction relevance of MultiModal and ReadWrite learning styles than the others (See Table 3). Also, the range of F-measure (0.765–0.875) of MultiModal and ReadWrite learning styles indicates a better accuracy of classification. Precision and Recall values of Visual and Auditory learning styles indicate moderate prediction relevance. Kinesthetic is poorly predicted and classified by all classifiers.

6. Conclusion

This study focused on assessing the resultant outcome of using competing perspectives to predict learning style. Learning styles provide clues about how learners see, engage with, and respond to their learning situations. Several algorithms have been used to assess the learner’s preferred method of learning using a specific learning style model. Using WEKA software, experiment was conducted using the Expectation—Maximization clustering and Naive Bayes, J48, OneR, and SMO classifiers to extract patterns in activity/event data to automatically identify learning styles of students. Data were validated and results were compared and affirmed. Each model correctly classified the training data in 0 second. The prediction relevance of MultiModal and ReadWrite learning styles and accuracy of classification were better than other learning styles. Kinesthetic was poorly predicted and classified by all classifiers. The automated technique is considered more reliable and superior as the results showed it performed better than the collaborative technique.

Results of the collaborative technique showed that multimodal (86, 46.7%) was the dominant learning style among the students, while visual learning style (17, 9.2%) was the least. The automated technique results showed visual, auditory, and read/write were higher than those identified from the questionnaire, with visual increased to 34, auditory to 41, read/write to 60, and multimodal reduced to 57. The experimental results indicated improved classification of learning styles and performance of students. This improved performance could have resulted from identifying the right learning styles of students and delivering appropriate learning resources. The students had different learning preferences in the same discipline, with multimodal and read/write learning styles being dominant. Easy understanding of textual learning objects and easy to access and understand varieties of learning objects also contributed to the improvement in learning outcome of the students. Students who understood varied and simple formats of learning imagery objects were more likely to respond effectively to related assessments, and those who received varieties of course contents and information had their performance improved.

7. Research contribution, limitations and future work

This study could serve as a base for further empirical investigation to gain in-depth knowledge in a quest to advance the ongoing research on technology-assistive learning and teaching techniques. Practically, practitioners would be able to manage wide-ranging datasets to respond swiftly to the requests of learners. We recommend extending this study to include diverse tertiary institutions, disciplines, stakeholders and both intrinsic and extrinsic factors to enhance findings. Also, we recommend complementing competing algorithms in future prediction.

Disclosure statement

No potential conflict of interest was reported by the author(s).