Predictive modelling and analytics of students’ grades using machine learning algorithms

Abstract

The outbreak of COVID-19 has caused significant disruption in all sectors and industries around the world. To tackle the spread of the novel coronavirus, the learning process and the modes of delivery had to be altered. Most courses are delivered traditionally with face-to-face or a blended approach through online learning platforms. In addition, researchers and educational specialists around the globe always had a keen interest in predicting a student’s performance based on the student’s information such as previous exam results obtained and experiences. With the upsurge in using online learning platforms, predicting the student’s performance by including their interactions such as discussion forums could be integrated to create a predictive model. The aims of the research are to provide a predictive model to forecast students’ performance (grade/engagement) and to analyse the effect of online learning platform’s features. The model created in this study made use of machine learning techniques to predict the final grade and engagement level of a learner. The quantitative approach for student’s data analysis and processing proved that the Random Forest classifier outperformed the others. An accuracy of 85% and 83% were recorded for grade and engagement prediction respectively with attributes related to student profile and interaction on a learning platform.

1 Introduction

With the Covid-19 pandemic putting the world in an unprecedented crisis, technology has played a vital role in maintaining continuity as far as possible (Dhawan, 2020). According to UNESCO, more than 990 million learners are affected by the crisis. The implementation of E-learning systems is the only sustainable solution. This change has caused multiple challenges and opportunities to the community that can be harnessed to improve the quality of service (UNESCO, 2020). Universities proved how E-Learning is beneficial for distance education.

The amount of data collected through E-learning platforms have massively increased over the last few years. As of now, more than 58 million students have registered for online courses worldwide with above 7000 courses offered (Moubayed et al., 2018). All the information collected can be utilised through machine learning and data analytics in the domain of online learning. Fields such as educational data mining and learning analytics are emerging with the aim to improve teaching/learning by making use of machine learning and visualisation techniques. Making efficient use of analysing and tracing data is still challenging. Machine learning is an efficient tool with the capacity to find hidden patterns of learner interaction. It can analyse complex non- linear relationships and has demonstrated to be a feasible approach in obtaining prediction of users on online learning platforms (Al-Shabandar et al., 2017).

Even though access is easier for developed countries, the dropout rate is higher compared to traditional modes of delivery. Consequently, assessing a student's performance is challenging. Academics are interested in forecasted results on assessments since they can direct their effort in improving the student’s experience (Bakki et al., 2015). At present, universities administer courses using online learning platforms. Analysts are making use of input features such as time, activity, assessment, and online discussion forums to forecast student performance. Recently, academics have been focusing a lot on predicting a learner’s performance and explaining the factors that affect the learning process (Sorour et al., 2015). The information collected can be used for decision making in terms of curriculum design, content, and mode of delivery. Students in universities are often unable to complete a course due to lack of understanding and engagement with the topics which is undoubtedly a matter of concern (Patil et al., 2018).

Being able to predict the grades of a learner is important in the learning process since it will help academics to understand the learner’s full potential and give the academics enough time to take corrective measures. Indeed the role of the academic should be to accompany the learner throughout his/her learning process and to be able to take corrective measures well before the exams. In line with what has been discussed above, the main aim of this research is to provide a predictive model to forecast students’ performance (grade/engagement) and to analyse the effect of online learning platform’s features. Implementing a working predictive software can be a baseline to initiate other research opportunities in the field of predictive analysis and the tools used for online learning. The wider implication of the study involves opening new avenues in terms of research in techno pedagogy. With predictive analysis in mind, academics can venture into new online learning tools, instructional design and adaptive learning techniques to revamp the content for students. The results of the analytics will be of assistance to students who are mostly likely to perform poorly. Corrective measures can therefore be implemented by academics and tutors.

1.1 Rationale and significance of research

Education plays an enormous role in what constitutes a society. Modern society is based on people who have high living standards and knowledge which allows them to implement solutions to challenging problems. Higher educational institutions are functioning under an increasingly convoluted environment. The competition among institutions, the response to local and global economic changes, politics and social changes are among a bunch of factors impacting the proportion of students, disciplines available and the overall quality. Institutions’ management are intended to adapt their decision-making process with the rapid changes occurring. Those decisions are often made without recourse to the vast data sources that are generated by manual and digital systems. The data, coupled with a predictive analysis system, can bring to light innovative action plans (Daniel, 2014).

To impart new skills and knowledge, universities have been making use of online learning platforms to deliver content. The global challenge for education is not just about providing access, but to ensure learning is taking place. To assess the comprehension of the subject, academics are making use of features such as online quizzes and discussion forums. Thereafter, students are examined through handwritten exams or assignments. The performance on the quizzes, discussion forum and the background of the students can be used to determine the grade for the handwritten exam or assignment. The performance allows the evaluation if a student has grasped the knowledge imparted and provides a scientific approach to investigate gaps (Yin, 2021).

1.2 Research questions

Research questions act as a catalyst for research projects and help to focus on the steps that will be taken to produce the analysis, findings, and results. Kitchenham's approach was considered to create the research questions. This approach takes into consideration the Population, Intervention, Context and Outcome (PICO) (Shahiri et al., 2015). The criteria were defined below in Table 1 below.

Table 1 Criteria for research questions

The research questions were framed in the Table 2 below.

Table 2 Research questions

2 Literature review

2.1 E-learning and online learning platform

E-learning is the delivery of education and all related activities using various electronic mediums such as the internet. It has provided several benefits like learner’s flexibility and an increase in interactions through both asynchronous and synchronous in the form of digital activities by using a learning management system (LMS) (Coman et al., 2020). Asynchronous e-learning is the most prevalent form of teaching/learning technique of E-learning due to its flexible methodology. LMS is used in an asynchronous environment to provide students with available learning objects in the form of video, document, audio, presentation and so on. The online learning platform provides the framework for students to view and communicate asynchronously and act as a repository for learning objects (Cohen & Nycz, 2006). Quizzes and assignments are also among the most helpful asynchronous activities for education (Perveen, 2016).

The use of LMS has increased tremendously in higher education. Discussion forum is a helpful asynchronous approach to initiate exchange of ideas and to participate on a particular topic. The students, at their own comfort, can contribute on the platform (Shida et al., 2019). Other tools such as quizzes can act as a self-assessment exercise to help students to improve understanding of concepts. LMS are usually able to capture student’s data and activities. One area of research is the multi-faceted benefits when exploiting the data (Shida et al., 2019). Devising asynchronous e-learning policies can increase student motivation, participation, problem solving, analytics and thinking skills (Adem et al., 2022; Chilukuri, 2020).

2.2 Student engagement

The interest in exploring learning analytics related to student engagement has been growing considerably lately. This has further expanded the research field for education. Higher education institutions have shown their interest in making use of analytics to support their engagement. This can act as an instrument that will help in mediating student/teacher information sharing resulting in effective learning, improve awareness, and a way to tackle current challenging situations (Silvola et al., 2021). Students who are engaged in their activities normally perform well and take pleasure in learning new content. Research has revealed that student engagement influences cumulative learning, long-term achievement, and promotes overall learner’s well-being (Salmela-Aro & Read, 2017). (Dewan et al., 2019) reviewed the engagement detection techniques in an online learning environment with its challenges. The detection methods were classified as automatic, semi-automatic and manual. Techniques in the automatic category obtain data from various sources. Log-files have been an efficient way of extracting valuable information especially in an online learning environment. (Cocea & Weibelzahl, 2011) analysed logs generated in an online platform known as HTML-tutor. They were able to extract 30 attributes such as number of tests attended, correct answers given number of pages accessed and so on (Dewan et al., 2019).

2.3 Machine learning

The purpose of machine learning is to obtain information from data which is why it is closely related to statistics, AI, and computer science. There are 3 types of machine learning techniques namely supervised learning, unsupervised learning, and reinforcement learning (Müller & Guido, 2017). Machine learning algorithms that learn from inputs and respective output pairs are known as supervised learning algorithms. Their purpose is to be able to generalize from known examples to automate decision-making processes (Müller & Guido, 2017). Though it is difficult to mount and analyse a dataset, supervised learning algorithms are popular, and their performance is easy to calculate. Unsupervised learning looks for undetected patterns in an unlabeled data set and little human supervision. In reinforcement learning, an agent will observe an environment to learn and achieve a goal. The computer employs trial and error to solve a problem (Russell, 2018).

The efficiency of a machine learning solution relies on the nature of the dataset and performance of the algorithms. Selecting a proper learning algorithm that is suitable for an application in a particular domain is strenuous. The reason behind this is that the purpose of ML algorithms is different. Even the outcome of different learning algorithms in a similar category may vary depending on the data characteristics (Sarker et al., 2019). Many machine learning algorithms have been implemented in the research community. Among the most important and famous techniques that figure in data science literature are listed below (Russell, 2018).

• 1. Logistic regression

• 2. K-nearest neighbors

• 3. Naïve Bayes

• 4. Decision trees

• 5. Random forests

• 6. Support vector machines

• 7. Deep Learning

2.4 Logistic regression

Basically, linear regression is performed to determine the relationships between two or more variables impacting each other, and to make predictions by making an analysis on the variations (Uyanık & Güler, 2013). Models requiring more than one independent variable are known as multiple linear models. The equation describes how independent variables affect the dependent variable (Petrovski et al., 2015).

$$\mathrm{Y}=\upbeta 0+\upbeta 1\ \mathrm{ x}1+\upbeta 2\mathrm{\ x}2+...+\mathrm{\beta n\ xn}$$

Whereby x is an independent variable, y is the dependent variable, Β1,β2,……βn are unknown parameters (coefficients) and β0 is a constant to create a line of best fit. Using linear regression is not convenient for categorical output. Considering for example a two-class classification problem, linear regression is prone to plot inaccurate decision boundaries in the presence of outliers. Logistic regression was developed for classification problems. The objective of logistic regression is to map a function from the features of the dataset to the targets to calculate the probability that a new entry belongs to one of the target classes (Bisong, 2019).

2.5 K-nearest neighbors

It is among the most basic and straightforward classification techniques. This method is suitable when there is little or no information about the distribution of the data. KNN was developed when reliable parameters to estimate probability were unknown or hard to establish (Hall et al., 2008). A parameter named k determines how many neighbors will be selected for the algorithm (Zhang, 2016). The performance is mainly determined by the choice of k and the distance metric used. If k is small, the estimate tends to be poor because of sparseness in data. Large values of k cause over-smoothing, performance degradation and miss out on important patterns (Zhang, 2016).

The aim is to choose a suitable k value to balance out overfitting and underfitting. Some researchers suggest setting k equal to the square root of the number of observations in the dataset (Gil-García & Pons-Porrata, 2006).

The k-nearest-neighbor classifier is commonly based on the Euclidean distance between a test sample and the specified training samples. By default, the knn() function use the Euclidean distance which can be determined with the equation:

$$D(p,q)=\sqrt{{\left({p}_{1}-{q}_{1}\right)}^{2}+{\left({p}_{2}-{q}_{2}\right)}^{2}+\cdots +{\left({p}_{n}-{q}_{n}\right)}^{2}}$$

Whereby D is the Euclidean distance and p and q are subjects to be compared with n characteristics (Zhang, 2016).

2.6 Naïve bayes

It makes use of a simple probabilistic function for classification. It computes a set of probabilities by calculating the frequency and combination of values in a dataset. It allows all attributes to contribute to the final decision equally. (Wibawa et al., 2019)

$$P\left(Q|X\right)=\frac{P\left(X|Q\right).P(Q)}{P(X)}$$

With

X:

Data with unknown class

Q:

The hypothesis X is a specific class

(Q|X):

The probability of the Q hypothesis refers to X

(Q):

Probability of the hypothesis Q (prior probability)

(X|Q):

Probability X in the hypothesis Q

(X):

Probability X

Naïve Bayes works well with high-dimensional sparse data and is insensitive to irrelevant data or noises. Its simplicity and low execution time makes it an ideal choice for predictive analysis. (Müller & Guido, 2017).

2.7 Decision trees

A decision tree has a tree-like structure where each node shows an attribute, each link shows a decision (rule) and each leaf shows an outcome. It can be used for both continuous and discrete data sets (Patel & Prajapati, 2018). Decision tree begins with a root node. From this node, users split each node recursively according to a decision tree learning algorithm based on if-the questions (Yadav & Pal, 2012). The result is a decision tree in which each branch represents a possible scenario of decision and its outcome (Sungkur & Maharaj, 2022). An example of a decision tree is shown in Fig. 1 below.

Fig. 1

Structure of decision tree algorithm (Hafeez et al., 2021)

2.8 Random forests

Random forest is considered as an expert solution for the majority of problems and falls under the ensemble learning classifiers whereby weak models are combined to create a powerful one. Ensemble methods are among the most promising areas for research. It is defined as a set of classifiers whose predictions are brought together to forecast new instances. Ensemble learning algorithms have shown to be an efficient technique to improve predictive accuracy and dampen learning problem complexities into sub-problems (Krawczyk et al., 2017). Numerous decision trees are produced in random forests. To classify an object having attributes, every one of the trees gives a classification which is also considered as a vote. The forest is then given the ability to choose the classification with the maximum votes. This is shown in Fig. 2 below.

Fig. 2

An example of a random forest structure considering multiple

2.9 Support vector machines

The basic idea of SVM is to plot data in n-dimensional space with n number of features and apply a hyperplane to distinguish the classes which are used for classification and regression (Deepa & Senthil, 2020). The input space is mapped to a high-dimensional feature based on a transformation defined by a kernel function (Ø) (Theobald, 2017). This is shown in Fig. 3 below.

Fig. 3

Transformation of data into a higher dimension with the kernel function (Theobald, 2017)

The hyperplane classifies the data separated by boundaries produced by the hyperplanes that separate classes of data points (Nayak et al., 2015). The optimization objective is to maximise the margin which is the distance between the separating hyperplane, decision boundary, and the training samples that are closest to this hyperplane (Raschka & Mirjalili, 2017). Using large margins tends to produce lower generalisation errors in models where a small margin is more likely to overfit. This is illustrated in the Fig. 4 below.

Fig. 4

Decision boundary distance in SVM (Raschka & Mirjalili, 2017)

2.10 Deep learning

Deep learning is a subfield of ANNs termed as such due to its use of multiple layered neural networks to process data. The idea is to have hidden layers (called hidden since they do not receive the raw data) combine the values in the preceding layer to learn the more complicated function of the input. (Sungkur & Maharaj, 2021) presents a research where ANN with Backpropagation Algorithm is used to provide personalised learning for cybersecurity professionals. This approach addresses the problem of ‘one-size-fits-all’ learning and makes the learning process more motivating, engaging and effective.

It is challenging for computers to understand raw data. This is where deep learning decomposes challenging problems into a series of nested concepts where each is described by a different layer of the predictive model (Di Franco & Santurro, 2020). Implementing typical machine learning algorithms is usually repetitive with lots of trial-and-error methods. Selecting different algorithms will produce different results which can be acceptable in several contexts. Nevertheless, with the limitations of different algorithms and the upsurge in machine learning theories and infrastructure, deep learning is, as technique, a more profound way of explaining high/low level of abstraction for a given dataset which typical machine learning algorithms are unable to do (Beysolow, 2017).

2.11 Machine Learning life cycle and methodology

Machine learning has its own life cycle that is the process the data undergo for the development and deployment of a predictive system. As compared to software development life cycle, the development of machine learning models involves experimenting on datasets to achieve the aims and objectives defined when applying fresh data after training (Ashmore et al., 2019). The basic workflow necessitates extraction of data, training, testing, tuning and evaluating the model before deploying it to production (Landset et al., 2015). This is shown in Fig. 5 below.

Fig. 5

Basic machine learning workflow (Landset et al., 2015)

Machine learning systems demand that data are in a certain format to be fed and processed. Essential processing activities are performed such as cleaning of unusual values, handling mistakes, formatting and normalisation.

2.12 Data sources & education database system

The data sources have to be identified for extraction and consolidation in a container to facilitate access. The education system is continuously expanding with an increase in the number of students. The student information has therefore increased considerably causing a lot of pressure in information organisations both at academic level (Yin, 2021). Research and academic institutes are working to uncover new theories from knowledge discovery. Modern educational institutions are constantly undergoing digital transformation both in terms of administrative and teaching/learning services. A comprehensive digitisation of the education process is at the root of all research. This will increase the attention of researchers in the field of data science and machine learning (Yin, 2021).

The core idea is to have the centralised database act as a data warehouse used to process and manage data (Jayashree & Priya, 2019). The data from multiple heterogeneous sources are put together in an organised and easily accessible manner. This enhances decision-making and provides greater insight in an organisation’s operation. Data warehousing, mining and analytics are famous in the business world. Its usage is still low in educational institutions. However, different studies and research areas in educational data mining are motivated to have their analytical processes applied to a database. The need of a data warehouse is obvious for learning analytics and evaluation of teaching–learning techniques (Moscoso-Zea et al., 2018). Quality-wise, it can be an instrument in obtaining organisational knowledge (Moscoso-Zea & Lujan-Mora, 2017).

Educational institutions with centralised databases can improve information management. Strategy implementation by board of directors, recruitment decision, retention and performance of students are some among several benefits that organisations can exploit (Williamson, 2018). Modern database system is emerging in the education domain leading to the emergence of Education Data Mining (EDM) as a field. It is playing a huge role in identifying patterns in learning principally performance. Predicting performance, success and retention rate with an e-learning environment as backdrop is becoming essential (Alyahyan & Düştegör, 2020). There are different approaches in building a database system/warehouse. Kimball's methodology uses a bottom-up approach which is convenient for projects having limited time and usually on a budget (Kimball & Ross, 2013). This is shown in Fig. 6 below.

Fig. 6

Ralph Kimball's bottom-up approach to DWH design. (Kimball & Ross, 2013)

A study by (Moscoso-Zea et al., 2016) suggested the design of Kimball’s to be convenient in educational institutions. The main reason is that their units are not integrated and usually function individually. Implementation design is challenging but beneficial for data consolidation and analysis. (Moscoso-Zea, et al., 2018). The major benefit of centralising the data is the possibility of having multiple client applications retrieving data simultaneously. All data stored one place allows easier querying and benefits in terms of execution time. Other applications, for example analytics software, can plug into the database (Singh, 2011).

2.12.1 Performance metrics

Evaluating the predictive model is an essential part to determine the accuracy of the student’s performance. To do so, it is important to quantify the quality of a system’s predictions (Mourdi et al., 2019). Some important performance metrics to assess the machine learning techniques are:

Accuracy

It is defined as the ratio of correct predictions to total number of sample input. It is a frequently use metric to assess the quality of a classifier’s solutions. It is the most used evaluation metric for both binary and multi-class classification. It is a determining value for assessing the capability of an algorithm

$$Accuracy=\frac{True\ Positives+True\ Negatives}{Total\ Number\ of\ Sample}$$

Precision

It is the number of correct positive results divided by all samples labelled as positive by the algorithm.

$$Precision=\frac{True\ Positives}{True\ Positives+False\ Positives}$$

Recall (Sensitivity)

It is the number of correct positive results divided by all samples that should have been labelled as positive by the algorithm.

$$Recall=\frac{True\ Positives}{True\ Positives+False\ Negatives}$$

F-measure (F₁-score)

A model can have a high recall value with low precision. Those values alone are not enough for indicating a good classifier. F-measure represents a harmonic mean of precision and recall. A higher value designates a high classification performance (Table 3).

Table 3 Variables for performance metrics (Michelucci, 2019)

$$F-measure=\frac{2\ \times \ precision\ \times\ recall}{precision+recall}$$

The variables used in the equation are defined as follows:

2.13 Related works

Considerable research has been conducted to forecast student performance. Researchers have gone through different methodology to showcase results to support their findings in terms of evaluation metrics. Lately, (Adnan et al., 2021) investigated the capabilities of seven algorithms such as the traditional SVM, KNN, ensemble techniques and deep learning. Though having recorded accuracy close to 91% with random forest, the attributes didn’t cater for several online learning tools. The student profile and number of clicks are among the 13 attributes used for prediction on a large dataset of 35,593 imbalanced records. (Ko & Leu, 2021) happened to record 82.26% with Naïve Bayes on a small dataset of 215 students without any balancing technique. The latter did not include online learning features as attributes. The e-learning aspects have been considered by (Mourdi et al., 2019) together with a dataset of 3585 students. The 25 attributes include information from quizzes, videos and forums. Though unbalanced, an accuracy as high as 99% was obtained for identifying a student as pass, fail or drop out. (Bujang et al., 2021) and (Costa et al., 2017) catered for imbalance classes during their analysis. (Bujang et al., 2021) indicated how the combination of SMOTE and feature selection influence accuracy of predictive models. (Costa et al., 2017) coupled SMOTE with fine tuning of algorithms. However, no comprehensive documentation was mentioned in terms of hyperparameters and values. (Tarik et al, 2021) opted to remove all missing data from its initial 142,110 students. With the remaining 72,010, accuracy of up to 70% were recorded with Random Forest.

Liu et al. (2022a) highlights on the importance of emotional and cognitive engagement as two prominent aspects of learning engagement. The authors further discuss about how emotional and cognitive engagement further share an interactive relationship and that these two factors thereafter jointly influence learning achievement. Liu et al. (2019) presents an unsupervised model, namely temporal emotion-aspect model (TEAM), modelling time jointly with emotions and aspects to capture emotion-aspect evolutions over time. Liu et al. (2022b) explores the relationship between social interaction, cognitive processing and learning achievement in a MOOC discussion forum. Liu et al. (2022c) discusses the relationship between discussion pacing (i.e., instructor-paced or learner-paced discussion), cognitive presence, and learning achievements. Emotion experiences, cognitive presence or social interactions in discourses as highlighted by the works of (Liu et al., 2019, 2022a, 2022b, 2022c) provide some deeper and implicit features that have a definite impact on the learning achievement of the learner.

3 Proposed solution

3.1 Research design

Experimental research is essentially the investigation of one or more variables (dependent variables) manipulated to assess the effect on one or more variables known as independent variables. It is based on the cause-and-effect relationship on a chosen subject matter to conclude the different relationships that a product, theory, or idea can produce (Jongbo, 2014). The nature among the variables is established with precise and systematic manipulation. This technique is suitable where testing theories and evaluation of methods are at the core of a study. Furthermore, the same set up and protocol can be replicated with the same variables. This can substantiate the validity of products, ideas, and theories. (Wabwoba & Ikoha, 2011). Additionally, this type of scientific approach can provide a set a guideline for evaluating and reporting information for research (Marczyk et al., 2005). Figure 7 below shows a popular general aspect of how experiments are conducted before reaching model evaluation.

Fig. 7

General Research Approach for machine learning (Kamiri & Mariga, 2021)

To empirically assess the algorithms and interpret the research outcome, the criteria used for the experimental procedures will be set as:

• The algorithms running

  • Type of supervised learning algorithm

• The evaluation technique

  • The training and testing procedures (E.g Cross validation)

• Predictive performance on unseen data

  • This involves estimation metrics such as percentage accuracy.

• Model specific properties

  • Hyperparameters (E.g. depth of a decision tree)

In this research, the controlled experiment will be related to the machine learning algorithms used and the estimation of each model on unseen data. It will be a convenient method to discover the techniques chosen that work best for the dataset and under which specific conditions with systematic experimentation. A comparative model analysis will have the supervised learning algorithms as variable and the assessment criteria as dependent variable. This is shown in Fig. 8 below.

Fig. 8

Research design derived from research questions

3.2 Machine learning architectural design

To interpret the processes, a new workflow with feature selection techniques and a way to handle imbalance classes have been incorporated in typical ML procedures (Fig. 9). It is represented in the architecture below.

Fig. 9

Machine learning architectural design to evaluate student performance

3.3 Web scraper

Due to the restriction on accessing the LMS database and limitations, a web scraper is required to retrieve data pertaining to discussion forums by sifting through the web pages of the internet-facing application. A web scraper was developed as shown in Fig. 10 below.

Fig. 10

Web Scraper Architecture for retrieve discussion forum information

3.4 Data consolidation & database

The data from the LMS, examination and student section have to be consolidated into a database. It is advantageous to have all data under one umbrella prepared for data extraction. A csv file can then be generated. The tabular dataset will be fed to the models (Fig. 11).

Fig. 11

CSV format generated from database

3.5 Proposed framework

The practices underlying the concept of data consolidation, processing and evaluation of student performance were translated into a framework. The same concept can be re-applied in different educational contexts (Fig. 12).

Fig. 12

Proposed framework for predicting student performance and engagement

3.6 Dataset

The data digitally available was extracted and cleaned. As per the cohorts available, a total of 1074 students’ data was used (Table 4).

Table 4 Dataset after feature encoding for predicting student’s grade

3.7 Student performance prediction software

The implementation of a software for predicting performance will address challenges in a systematic manner. The functional requirements describe the intended function of the Student Performance Prediction Software and are shown in Table 5 below.

Table 5 Functional and non-functional requirements

A web application was developed to execute the machine learning algorithms as per the best instance. The rationale behind it is to have a front-end web interface where users will be able to upload a file with a student’s data. The interface will then predict the student’s performance and engagement. The application can be deployed on a production environment for the institution (Fig. 13).

Fig. 13

Web application for student performance and engagement prediction

4 Results and discussions

4.1 Testing and evaluation

The accuracy of a model is the primary indicative factor to assess a model. The results obtained per algorithm were reported according to the highest accuracy observed. For the same configuration providing the best accuracy, the average total precision, recall and F1-scores per fold were recorded and illustrated in the subsection. An average of the confusion matrix per fold was estimated. The machine learning algorithms that were compared and contrasted include Logistic regression, K-nearest neighbors, Naïve Bayes, Decision trees, Random forests, Support vector machines and Deep Learning. It was observed that Random Forests yielded the best results. For the sake of simplicity, the diagrams of only Random Forests are shown below.

4.1.1 Random forests—Student grade prediction

Figure 14

Fig. 14

Confusion matrix for grade prediction using RF

Figure 15

Fig. 15

Average of Evaluation metrics per fold—grade prediction using RF

Table 6

Table 6 Performance metrics for grade prediction using RF

4.1.2 Student engagement prediction

Figure 16

Fig. 16

Confusion matrix for engagement prediction using RF

Figure 17

Fig. 17

Average of Evaluation metrics per fold—engagement prediction using RF

Table 7

Table 7 Performance metrics for engagement prediction using RF

Following implementation and evaluation stages, all the functional and non-functional requirements have been achieved. The results were studied to answer the research questions initially set.

4.2 Research questions

This section discusses the answers to the research questions set earlier. This further helps to shed light on how machine learning algorithms can be used for predicting students’ grades.

RQ 1. How precise are the machine learning algorithms at predicting students’ performance (Grade & engagement)?

Figure 18

Fig. 18

Evaluation of ML Models for grade prediction

Figure 19

Fig. 19

Evaluation of ML Models for engagement prediction

Evidence revealed Random Forest outperformed its counterparts in accuracy, prediction, recall and F1-score both for predicting grade and engagement level. The algorithm’s properties seem to be effective for classification of such a peculiar dataset with the application of MICE as imputation technique, feature selection, SMOTE and normalisation. RF is particularly advantageous when dealing with high dimensional attributes. Setting a high number of trees (n_estimators = 1000 and 100 trees in forest) as per the number of attributes unveiled an accuracy of 85% and 83% for both models respectively. The evaluation metrics for all classifiers oscillate in every fold. However, for the engagement prediction, MLP suffered a drastic drop in the second fold. It may have been exposed to data which is beyond its training configuration. Anomalies as such hurt a model as it is preferable to have a high metric value across all folds to ensure generalisation (Fig. 20).

Fig. 20

Drop in 2nd Fold for Average of Evaluation metrics per fold—engagement prediction using MLP

Nevertheless, RF, SVM, DT, KNN & MLP obtained above 70% as average in all metrics. It can be concurred that the classifiers are applicable in an education-related context with multiple student attributes, both personal data and interaction in an online environment for grade/engagement classification.

RQ 2. What are the important attributes in predicting the students’ grade?

Initially 42 features were identified for processing. Following imputation and category encoding, the feature selection technique discarded the redundant attributes that would not benefit the model. After MICE imputation, 16 features were removed. Collecting data can be expensive and since the redundant attributes can be ignored, the focus can be shifted on other new features. Among the 30 remaining attributes, 9 out of 10 learning objects were retained. A deeper understanding of the discussion forums and the way students tackle the MCQs can open promising directions. Only the total number of discussions, participation and the time taken to submit and MCQ was utilised. This can be a reference guide for potential research with respect to interactions and duration on a platform.

RQ 3. Can an adaptable predictive modelling framework be developed for student performance and engagement?

The architecture can accommodate new data and techniques for experimentation such as including new learning objects and going through different ML processes. The application can be deployed in different contexts as per the requirements and available data on or off premises. The prospect for new E-learning strategies and predictive analytics are enormous. The framework can eventually be transposed and adapted in different educational establishments (primary school, secondary school, training centres etc.) for experimentation and at production level. Having analytic tools incorporated in the educational system will allow institutions to provide a level playing field for scholars. The study can bring to the fore the importance of digital transformation and unification of data sources in a single container for analysis which is essential for E-learning analytics. The major achievement of this research and the difficulties encountered are outlined below.

4.2.1 Major achievements

• 1. A machine learning architecture and comprehensive life cycle was set up. The system implements all the machine learning phases such as imputation, feature selection, class balancing etc.…

• 2. The predictive framework for student grade and engagement level was translated into a working prototype through a web application. An engagement threshold was estimated.

• 3. The hyperparameter tuning was documented together with values and parameters.

• 4. The system was reworked to approximate confusion matrices and evaluation metrics per fold. A stratified cross validation of 10 folds were imposed and the results provide an average of the confusion matrix most likely to be generated from a set of data. The confusion matrix and charts are indicative with respect to the overall performance of classifiers.

• 5. Data consolidation and cleaning is an intricate process with hiccups at every turn. A web scraper has been developed to circumvent data retrieval issues which is common practice in data science projects. The centralised database was set up to assemble all relevant data.

• 6. Using latest releases of ML libraries (E.g., Keras) and resorting to release documents for guidance during development.

4.2.2 Difficulties encountered

• 1. The databases are not centralised. There are no unique identifiers present across all data sources making it difficult to extract, compare, and process data.

• 2. The log files are large causing the download process to often time out. The configuration setting on the server cannot be altered neither the settings on the web application. The log files also generate fields that are irrelevant to the. The log files were generated and downloaded in stages to avoid retrieving large files. The logs were then cleaned, and only pertinent data were retained.

• 3. Processing and debugging are tedious due to the execution time of the algorithm’s constraint by the hardware.

5 Conclusion

This research can contribute to create actionable steps for growth to improve educational institutions’ reputations and ranking both at national and international level. The software will be at the disposal of experts and will act as a device to help in reinforcing the learning process for existing or novel pedagogical interventions. Applying the framework as a magnifying glass on the education system can make way for innovative concepts that will undoubtedly bring waves of change in the learning process. From a scientific point of view, every phase of the machine learning life cycle can be further explored. Data scientists now researching new filtering algorithms, imputation techniques and normalisation procedures can measure their efficacy in the education context. Random Forest classifier outperformed the other classifiers. An accuracy of 85% and 83% were recorded for grade and engagement prediction respectively with attributes related to student profile and interaction on a learning platform. From an educational point of view, this research can help educators identify learners that are at risk as far as poor performance is concerned and can help the educators take timely corrective measures.

One of the limitations of this research is that external factors might be affecting the student performance when participating in discussion forums and quizzes, for example, the bandwidth and performance of computer or mobile devices might be impacting on the participation of the learner in certain learning activities. Future studies can include the investigation and implementation of strategies in context with modern learning techniques such as personalised learning for isolated learners. Moreover, to supplement the quantitative approach, qualitative research methods can be combined to gather insights about the learning process and outcome of students. For example, analysing the students’ feedback can be included in the existing experimental setup, subject matter experts for a given material can examine students’ response and uncover other areas in teaching and learning when preparing the dataset. Future works also include multi-feature fusion since it would be interesting to feature out the causal relationship of emotion, cognition, behaviours and motivation behind learning performance and how to further improve it.