Evolutionary Algorithm Based Feature Subset Selection for Students Academic Performance Analysis

1  Introduction

Recently, different kinds of learning management systems (LMSs) have been effectively adopted by higher education institutions and universities, collecting large amounts of educational data and recording various student learning features [1]. Educational Data Mining (EDM) is a rapidly expanding scientific field that provides the potential to analyze these information and harness useful knowledge from it. Eventually, a plethora of prediction algorithms have been efficiently used in an educational context to resolve a great number of challenges [2]. But, creating prediction methods in the field of EDM by using TL models has been poorly studied until now. Thus, the most important question in current research is whether a prediction algorithm trained on a previous course would perform well on a novel one [3]. observes that a course (a) is populated with distinct instructors and students, (b) may have features that could not be transmitted (for example, a feature determined on specific learning resources that is inaccessible on other courses) and (c) would change over time in different ways, despite being structurally and contextually distinct. Additionally, the difficulty in course design and the LMS have a great influence on the course progress at the time of semester. Hence, there might be a problem where the transfer learning (TL) method might not reflect the anticipated result, which shows some uncertainty regarding the prediction accuracy of the recently established learning algorithm [4].

Prediction and students’ performance analysis are the two extensively discussed areas of research in educational systems [5]. Despite their disparate goals, the outcomes of performance analysis have a significant impact on predictive research. The analysis of student progress during their studies offers university management data regarding the possibilities of success for all the students. Conventionally, this analysis can be made by the lecturers who utilize their communication with students in classroom activities and mid-term assessments to take timely action and find those “at risk” of dropping out [6]. In the current scheme of higher education, the communication time between students and lecturers is continuously decreasing and finding endangered students has become even more challenging. This is because of the increasing number of students and access to online learning resources for students [7]. More precise predictions of student’s success could be achieved by the current systems of machine learning and data mining analysis.

The prediction of learner performance can be accomplished with the help of several data mining approaches. For example, the machine learning (ML) methods as stated in research works of [8] 2020. The data mining systems are categorised as (a) ML algorithms (like neural networks, symbolic learning, swarm optimization), (b) statistical approaches (namely cluster analysis, regression analysis, discriminant analysis) and (c) artificial intelligence techniques (for example, fuzzy logic, genetic algorithms, neural computation). In this regard, the present study aims at examining the capacity of artificial neural networks (ANN) to forecast team’s performance. Generally, ANN is a subdivision of Artificial Intelligence (AI), which includes the application of Deep Learning (DL) and Machine Learning (ML). The ANN theory was stimulated by the human brain and especially by the biological neural networks. It contains synapsis node for perceptrons and artificial neurons.

The synopsis is found in neural connections and the biological brain is used for learning and memory. The connection strength, i.e., the synaptic weights corresponding to the memorized data, can be changed at the time of the learning method. An array of neurons composes an ANN layer [9]. Usually, all the neurons receive signals (the weighted amount of their inputs), next process them and employ an activation function for signaling neurons related. A simple ANN framework includes output, input, and hidden layers. In contrast, a Deep Neural Network (DNN) method consists of (a) the output layer, (b) the input layer and (c) more than two hidden layers (Denses). The basic functions for training and activating the DNNs are the loss function (feed-forward loop), the optimizer (back-propagation loop) and the activation function [10].

This paper designs a novel metaheuristic optimization-based feature subset selection with an optimal deep learning model (MOFSS-ODL) to predict students’ performance. The proposed MOFSS-ODL technique uses an isolation forest (IF) based outlier detection approach to eliminate the existence of outliers. Moreover, the Chaotic Monarch Butterfly Optimization Algorithm (CBOA) is applied to choose highly related features with low complexity and high performance. Furthermore, a sailfish optimizer with stacked sparse autoencoder (SFO-SSAE) approach is employed to classify the educational data. To validate the enhanced predictive outcome of the MOFSS-ODL technique, a series of simulations were carried out against a benchmark students’ performance data set from the UCI repository and the results are inspected under several dimensions.

2  Related Works

This section offers a comprehensive review of existing students’ performance analysis prediction models. In Akour et al. [11], an effort has been made to examine the efficacy of utilizing the DL method more precisely to forecast students’ achievements, which could assist in forecasting when the student can complete their degree or not. The simulation result shows that the presented method outperforms the current approach in terms of predictive performance. In Harvey et al. [12], a prediction method is examined and developed in this work for analysing a K-12 education data set. The LR, DT and NB methods are among the two classifications used in the development of these prediction methods. The NB method gives the most accurate predictions for high school students’ SAT and Math scores.

Orji et al. [13] carried out prior research on determining methods of improving student engagement in online learning techniques via data-driven intervention. Student engagement in this work can be described by objective information (activity log of a certain UG course in a TELS). Activity logs are unbiased data and a reflection of students’ actual learning behaviour (uncontrolled). In this work, the logs of student learning activities from TELS are mined, which employed UG courses, to explore variances among student’s learning behaviours as they are related to their engagement levels and academic performance (evaluated based on last grade point on a course). Supervised (RF) and unsupervised (Clustering) ML methods are applied in examining the relationships.

Xu et al. [14] designed a new ML technique to predict students’ performance in degree programs, i.e., one capable of addressing these major problems. The presented model consists of 2 key characteristics. Initially, a bilayer framework including a cascade of ensemble predictors and multiple base predictors was proposed to make predictions according to the student’s evolving performance state. Next, a data-driven technique based on probabilistic matrix factorization and latent factor model is presented for discovering the course significance, i.e., significant enough to construct an effective base predictor.

Hamoud et al. [15] present method-based DT models and recommend the best method-based performances. This method makes use of the three created categories (REPTree, J48 Random Tree and the questionary filled out by the student). The study includes sixty questions that cover the subject, such as social activity, health, academic achievement, and relationships and how this affects student efficiency. A total of 161 questionnaires were collected. This approach was developed using the Weka 3.8 tool [16] investigated data logged by a TEL scheme called Digital Electronics Education and Design Suite (DEEDS) using a machine learning technique. The ANN, SVM, LR, NB and DT classifiers are included in the ML method. When entering input data, the DEEDS approach allows students to perform digital design problems with varying levels of complexity. Reshma et al. [17] use a deep ANN on a set of hand-crafted features extracted from the virtual learning environment clickstream data to forecast at-risk students and provide a measure for earlier intervention in this situation. The results show that the provided strategy performs better in categorization.

Waheed et al. [18] developed a comprehensive study of limitations, hardware resources and performance. This study can be implemented by using Collaboratory to accelerate the DL method for CB and another GPU-centric application. The selected test-case is a parallel tree-based combinatorial search and 2 CV applications: object segmentation or localization and object classification or detection. The hardware under the accelerated running time is related to a conventional workstation and strong Linux servers armed with twenty physical cores. Cyril et al. [19] proposed a new end-to-end DL technique and suggested a DPCNN method for predicting students’ performances. This paper also presents multitasking learning methods and forecast the performance of students from different majors in an unified architecture.

3  The Proposed Model

In this study, an effective MOFSS-ODL technique was designed to predict students’ performance. The proposed MOFSS-ODL technique encompasses IF-based outlier detection, CBOA-based feature selection, SSAE-based classification, and SFO-based parameter optimization. Fig. 1 illustrates the overall working process of the proposed MOFSS-ODL technique.

Figure 1: Overall process of MOFSS-ODL technique

3.1 Isolation Forest Based Outlier Detection

At the primary stage, the IF technique is applied to eradicate the presence of outliers in educational data. The IF is an unsupervised algorithm employed in collective-based models to separate the anomalies by evaluating the isolation scores for each data point. The IF has a similar idea of utilising the tree algorithm as the RF model. It processes data points to recurrent random splits that are based on feature selection [20]. The major benefit of the IF technique is how, it processes information. Rather than processing each data point, it employs a DT mode for isolating the outliers, which decreases the processing and execution time and the respective memory requirements. The IF method functions by splitting the algorithm into various segments, which are needed for the subsampling size. The IF estimation starts with specific data points. Next, based on the selected value, it sets a range between the minimum and maximum values to define the outlier scores for all the data points in the tree. The score can be measured by setting a path length for isolating the outliers.

3.2 CBOA Based Feature Selection Technique

During the feature selection process, the CBOA is utilized to choose an optimal subset of features from the educational data; the concept of the BOA method is established in [21]. In BOA, each monarch butterfly is idealized and situated in Mexico (Land 2), southern Canada (Land 1), and the northern US. Next, the position of the monarch butterfly is upgraded in two ways, such as through butterfly adjusting and migration operators. Initially, the offspring are produced (a position upgrade) via the migration operators. Next, the position of another monarch butterfly is upgraded using the butterfly adjusting operators. Furthermore, both operators could be performed concurrently. Thus, the BOA method is appropriate for similar processing and has a good balance of diversification and strengthening. The BOA method includes two significant operators, which are listed below.

In migration operator, the aim is to upgrade the migration of monarch butterflies among Land1 and Land2. NP is the overall amount of monarch butterflies and NP1=cei1(p×NP) and NP2=NP−NP1  are the number of monarch butterflies in Land1 and Land2, respectively, whereas p represents the migration rate of monarch butterfly with p=5/12 in BOA, ceil (x) rounds x to the near integer better than or equivalent to x, the sub-population of Land1 is represented as Sub-population1 and the sub-population of Land2 is represented as Sub-population2. Next, the migration operators are denoted by

xi,kt+1={xr2,k′txr1,k′t r>pr≤p,   (1)

In which xi,kt+1 represent the kth component of xi in generation t+1 ; likewise, xr1,kt signifies the kth component of χr1 in generation t and xr2,kt indicates the kth component of χr2 in generation t; the present generation value is t, r1 & r2 Monarch butterflies are arbitrarily chosen from Sub-population1 and Sub-population2, respectively. Now, r is estimated by r=rand×peri , in which peri denotes the migration period, i.e., equivalent to 1. 2 in BOA and rand indicates an arbitrary value in the range of zero and one [22]. Fig. 2 illustrates the flowchart of BOA.

Figure 2: Flowchart of BOA

In the butterfly adjusting operator, it upgrades the location of monarch butterfly in Subpopulation2 as:

xj,kt+1={xbest,ktxr3,ktxi,kt+1+α×(dxk−0.5) rand ≤prand>pquadrand≤BARrand >pquadrand>BAR,   (2)

Let xj,kt+1 be the kth component of xj in generation t+1 ; likewise xbest,kt indicates the kth component of xbest in generation t, i.e., the optimal position for Land1 and Land2 monarch butterflies, xr3,kt signifies the kth component of xr3 in generation t, the monarch butterfly r3 is arbitrarily chosen from Sub-population2, also BAR indicates the adjustment rate. When BAR is lesser when compared to the rand arbitrary value, the kth component of xj at t+1 is upgraded, in which α indicates the weight factor and α=Smax/t2 , Let Smax  be the maximal walk steps. In (2), dx represents the walk steps of j butterfly which are evaluated by the Levy fight thus x=Leνy(xjt) .

The presented CBOA method depends on the incorporation of chaotic maps in the typical BOA. The major phases of the presented model can be given in the following.

a)   Invoke the chaotic map to update butterfly locations rather than utilizing arbitrary parameters hence it will enhance the CBOA accuracy. They would be adapted by substituting r2 with Cj as:

xit+1=xit+(Cj×g∗−xit)×fi (3)

xit+1=xit+(Cj×xjt−xkt)×fi (4)

whereas Cj represents the chaotic map and j=1, 2, …, 10 . Noted that the Cj value is a chaotic value, created by ten chaotic maps are substituted by r values for getting optimal outcomes in accuracy and minimal fitness when compared to original algorithm which uses arbitrary value.

b)   Transfer CBOA to binary CBOA: Employing the binary CBOA would characterise the searching space in binary values. Hence, the binary CBOA could adoptively search the feature space for optimal feature integration and is predicted to be very simple when compared to the continuous version. Binary CBOA can be given as follows:

xit+1={1 if (s(xit+1))≥rand()0 otherwise  (5)

Let s be a transfer function, rand () represents an arbitrary value made from uniform distribution [0,1] and xit+1 indicates the upgraded solutions.

s(xit+1)=11+exp10(xit+1−0.5)  (6)

Different from the standard BOA, wherever the solution was upgraded from the exploring space nearby to a continuing value place, in the CBOA, the searching space was demonstrated as a dimensional Boolean lattice. Besides, the solution was upgraded on the corner of the hypercube. In addition, to solve this issue of selecting or not, a given parameter and binary solution vectors were implemented; one relates to the parameter being chosen for comprising the new dataset, and zero relates to anything else. In binary techniques, one uses the step vector for evaluating the possibility of altering a place. The transmission function significantly influences the balance between exploitation and exploration. During the FS technique, if the size of the feature vector was large, the count of various features combined would inclines, for instance, a huge space for exhaustive research. The presented hybrid technique was utilized to achieve this resolve to explore the feature space vigorously and make the right combination of features. The FS falls into multi-objective issues as it requires satisfying many objectives to receive better solutions that minimize the subset of the FS and, concurrently, maximize the accuracy of the outcome for providing a classifier.

Based on what came before, the fitness function (FF) for finding solutions in this state was made so that the two goals could be met in a balanced way.

fitness =αΔR(D)+β|Y||T| (7)

ΔR(D) refers the classifier error rate. |Y| stands for the size of subset which this approach chooses and |T| entire number of features involved from the existing datasets. α defines the parameter ∈[0,  1] comparing with the weight of error rate of classification respectively but β=1−α signifies the significance of decrease feature. The classification efficacy was allowed a vital weight rather than the number of chosen features. If the evaluation function is only regarded as the classification accuracy, the outcomes are neglected of solutions that are comprised of the same accuracy but minimally chosen features that serve as essential aspects of decreasing the dimensionality issue.

3.3 SFO-SSAE Based Classification Model

Finally, the SFO-SSAE based classification model gets executed to determine the proper class labels for the educational data. The building block of deep networks for unsupervised learning features is a single AE [23]. It is another kind of ANN model and is architecturally determined by the output, input and hidden layers that compose a decoder and an encoder. In, the encoder transmits the data matrix to a hidden depiction with a tunable number of neural units, followed by a non-linear activation. The procedure can be expressed by Eq. (8).

h=f(x)=σ(W1Tx+b1)  (8)

In which, σ(⋅) indicates the activation function, W1 represents a weight matrix and b1 denotes a bias vector. Let Relu, sigmoid andtanh  be widely employed. During the decoder process, the deterministic mapping g(⋅) maps the hidden depiction back to recreate the x′ input space as:

x′=g(h)=σ(W2Th+b2) (9)

Whereas b2 & WV2 represents the bias vector and weight matrix of the decoder. The trained AE aim is to discover the θ=W1, W2, b1, b2 parameters which minimize the recreation loss among the output data X′ and input data X. The objective function is determined by:

arg minθ⁡JAE(θ) =1N∑i=1N⁡‖xi−xi′‖2+λ(‖W1‖2+W2‖2) (10)

Let λ be a standardization coefficient of the weight decay term, i.e., added to avoid over-fitting. The constraint representative power of a single AE results from the simplest shallow structural characteristics. Stimulated by the biological method of the human visual cortex, the conventional SAE can be generated by stacking AE consecutively, taking the output of the hidden unit of the prior layers as the input to the upper layers. Fig. 3 displays the framework of SAE. AE was initially presented according to the concept of reduction dimension and would lose the capacity to manually learn features if there were further hidden layer neurons when compared to input neurons. An SAE can be attained by adding a few sparse limitations to the conventional AE that could suppress all the outputs of hidden layer nodes. To attain this, the training objective of SAE can be expressed as follows:

arg minθ⁡∫SAE ⁡(θ) =1N∑i=1N⁡‖xi−xi′‖2+λ(‖W1‖2+‖W2‖2)+β∑j=1d⁡KL(ρ||ρ^j)  (11)

Here ρ^j represents the average activation values of each trained instance on the jth hidden nodes and it is limited by KL divergence to be closer to the sparse variable ρ.β control the weights of sparse penalty. Next, an SSAE is attained by presenting sparse limitations on every single AE of the SAE.

Figure 3: SAE structure

The SFO algorithm is utilized to determine the hyperparameters involved in the SSAE model. The SFO is a new, nature inspired meta heuristic approach, i.e., modelled after a group of hunting sailfish. It displays more competitive performances than widespread meta-heuristic models. The summary of SFO is demonstrated in the following. During the SFO method, the sailfish are considered as candidate solutions and the position of the sailfish in the searching space represents the variable of the problems. The location of ith sailfishes in kth searching iteration is represented as SFi,k and its equivalent fitness can be evaluated using f(SFi,k) . Sardine is the other important participant in the SFO method. Also, it is considered that a school of sardine is shifting in the searching space. The location of ith sardines is represented as Si and its equivalent fitness can be evaluated using by f(Si) . During SFO model, the sailfish occupying the optimal location is chosen as an elite sailfish that affects acceleration and maneuverability of the sardine at the time of attack. In addition, the location of injured sardines in all the iterations is chosen as an optimal location for collective hunting by the sailfish. The presented method’s aim is to avoid formerly discarded solutions from being elected again. Injured sardines and Elite sailfish are upgraded according to the Eq. (12).

YnewSFi=YeliteSFi−λi×(random(0, 1)×(YeliτeSFi−YinjuredSi2)−YcurrentSFi),  (12)

In which YcurrenTSFi represents the present location of sailfish and arbitrary value in the range of [0,1].

The λi variable represents the coefficients in the ith iteration and the values are acquired by (13):

λi=2×rand(0, 1)×SD−SD,   (13)

Let SD be the sardine density that represents the number of sardines in all the iterations. The SD variable can be acquired by Eq. (14):

SD=1−(NSFNSF+NS),   (14)

Whereas NS & NSF represents the number of sardines and sailfish, respectively. During the hunting process, sardines is injured or not tired and sailfish are energetic. Sardines could rapidly escape. But, with continuous hunting, the strength of sailfish attacks would progressively decrease. In the meantime, sardines would become tired and the alertness of the position of sailfish would also decrease. Consequently, the sardines are hunted. According to the algorithmic procedure, the novel location of sardines YnewSi is upgraded according to Eq. (15):

YnewSi=random(0, 1)×(YeliteSFi−YoldSi+ATP),   (15)

Now, YoldSi represents the older location of the sardine and arbitrarily value in the range of [0,1]. ATP signifies the sailfish attack power. The ATP variable is expressed by Eq. (16):

ATP=B×(1−(2×ltr×ϵ))  (16)

Here, B & ϵ represents coefficient that is utilized for reducing the attack power linearly from B to 0 as well as Itr indicates the iteration count. Since the attack power of sailfish reduces over the hunting time [24], this reduction promotes the convergence of the search. Once ATP is greater, e.g., larger than 0.5, the location of each sardine is upgraded. On the other hand, α sardines with variables β upgraders their locations. The number of sardines upgrades their location is defined by Eq. (17):

α=NS×ATP,   (17)

In the equation, NS represents the number of sardines in all the iterations. The number of parameters of the sardine which upgrader their locations is attained by Eq. (18)

β=di×ATP,   (18)

While di represents the number of parameters in the ith iteration. Once sardines are hunted, its fitness should be greater than the sailfish. In this condition, the location of sailfish YSFi is upgraded by the newest location of the hunted sardine YSi for promoting the hunting of novel sardines:

YSFi=YSi if f(Si)<f(SFi).  (19)

4  Experimental Validation

The benchmark dataset from the UCI repository is used to test the MOFSS-ODL model’s performance. The dataset comprises 649 instances with 33 features and two classes. Table 1 demonstrates the details of dataset decryption. Fig. 4 shows the frequency distribution of the attributes involved in the dataset.

Fig. 5 Investigate the best cost analysis of the CBOA with other FS models. The results showed that the information gain and CFS techniques have obtained poor performance, with the best costs of 0.38692 and 0.36600. At the same time, the PSO and GA techniques have attained moderate best costs of 0.18364 and 0.16528, respectively. The CBOA method, on the other hand, has led to better performance at the minimum best cost of 0.02411.

Figure 5: FS analysis of CBOA technique

The confusion matrices generated by the MOFSS-ODL model on the applied dataset under five distinct runs are in Fig. 6. The figures show that the MOFSS-ODL model has classified the instances into two classes effectively. For instance, with run-1, the MOFSS-ODL model has classified 538 instances into class 0 and 87 instances into class 1. At the same time, with run-2, the MOFSS-ODL method has classified 538 instances into class 0 and 86 instances into class 1. Also, with run-3, the MOFSS-ODL approach has classified 540 instances into class 0 and 86 instances into class 1. Concurrently, with run-4, the MOFSS-ODL technique has classified 538 instances into class 0 and 88 instances into class 1. Simultaneously, with run-5, the MOFSS-ODL system has classified 542 instances into class 0 and 88 instances into class 1.

Figure 6: Confusion matrix of MOFSS-ODL model with distinct runs

The values in the confusion matrices are transformed in the form of TP, TN, FP and FN in Table 2.

Table 3 offers a detailed classification result analysis of the MOFSS-ODL model under five distinct runs [26–29]. The experimental results reported that the MOFSS-ODL model has accomplished effective classification performance. For instance, with run-1, the MOFSS-ODL model has classified the instances with the precn , recal , Fmeasure , accuy and kappa of 0.9764, 0.9800, 0.9782, 0.9630 and 0.8231 respectively. In line with, with run-2, the MOFSS-ODL manner has classified the instances with the precn , recal , Fmeasure , accuy and kappa of 0.9746, 0.9800, 0.9773, 0.9615 and 0.8203 respectively. Moreover, with run-3, the MOFSS-ODL technique has classified the instances with the precn , recal , Fmeasure , accuy and kappa of 0.9747, 0.9836, 0.9791, 0.9646 and 0.8257 respectively. Furthermore, with run-4, the MOFSS-ODL approach has classified the instances with the precn , recal , Fmeasure , accuy and kappa of 0.9782, 0.9800, 0.9791, 0.9646 and 0.8259 respectively. Finally, with run-5, the MOFSS-ODL algorithm has classified the instances with the precn , recal , Fmeasure , accuy and kappa of 0.9783, 0.9872, 0.9828, 0.9707 and 0.8367 respectively.

Fig. 7 investigates the average classification outcomes analysis of the MOFSS-ODL model on the test dataset [30–35]. The figure depicted that the MOFSS-ODL methodology has resulted in maximal average precn , recal , Fmeasure , accuy and kappa of 0.9764, 0.9822, 0.9793, 0.96499 and 0.8263.

Figure 7: Average analysis of MOFSS-ODL model with different measures

Fig. 8 provides a comparative precn and recal analysis of the MOFSS-ODL model with recent methods. The figure portrayed that the RT and RRBFN models have obtained lower precn and recal values. In line with, the MLP and DT manners have reached to somewhat superior precn and recal values. Along with that, the LR, RF and ACO-LR models have resulted in considerable precn and recal values. However, the MOFSS-ODL model has gained improved outcome with the maximum precn and recal of 97.83% and 98.72% respectively.

Figure 8: Precision and Recall analysis of MOFSS-ODL model

Fig. 9 offers a comparative Fmeasure and Kapp analysis of the MOFSS-ODL manner with recent techniques. The figure outperformed that the RT and RRBFN manners have reached lesser Fmeasure and Kapp values. Besides, the MLP and DT models have reached slightly improved Fmeasure and Kapp values. At the same time, the LR, RF and ACO-LR models have resulted in considerable Fmeasure and Kapp values. However, the MOFSS-ODL model has reached higher outcome with the maximal Fmeasure and Kapp  of 98.28% and 82.63% respectively.

Figure 9: F-measure and Kappa analysis of MOFSS-ODL model

Accuracy analysis of the MOFSS-ODL technique with recent manners takes place in Fig. 10. The figure shows that the RBFN, MLP and RT techniques have obtained lower accuy of 89.52%, 89.67% and 87.67% respectively. Followed by the RF and DT models have attained slightly enhanced accuy of 92.60% and 91.37% respectively. Moreover, the ACO-LR and LR models have achieved moderate accuy of 94.91% and 94.76% respectively. However, the MOFSS-ODL model has resulted in a maximum accuy of 97.07%.

Figure 10: Accuracy analysis of MOFSS-ODL technique with existing approaches

After examining the tables and figures, it can be obvious that the MOFSS-ODL approach has accomplished effective outcomes on students’ performance analysis.

5  Conclusion

In this study, an effective MOFSS-ODL technique was designed to predict students’ performance. The proposed MOFSS-ODL technique encompasses IF-based outlier detection, CBOA-based feature selection, SSAE-based classification, and SFO-based parameter optimization. Besides, the utilization of CBOA-based feature selection and SFO-based parameter optimization help to accomplish improved students’ performance prediction outcomes. To validate the enhanced predictive outcome of the MOFSS-ODL technique, a series of simulations were implemented, and the outcomes are inspected in several dimensions. The experimental results pointed out the improved performance of the MOFSS-ODL technique over recent approaches in terms of several evaluation measures. In future, the predictive performance of the MOFSS-ODL technique will be enhanced by using the clustering approaches and hybrid algorithms for feature selection on educational datasets to predict students’ achievement.

Funding Statement: The authors received no specific funding for this study.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.