Arbitrator Miniature: A Paradigm using Data Science Methods to Predict Academic Performance

doi:10.21203/rs.3.rs-2271636/v1

Download PDF

Research Article

Arbitrator Miniature: A Paradigm using Data Science Methods to Predict Academic Performance

https://doi.org/10.21203/rs.3.rs-2271636/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Informational tools are necessary at schools and colleges due to the sheer volume and diversity of data they handle. Numerous scholars has emphasized towards applying machine learning to retrieve information from the education database to enable students and educators in attaining greater results as a means of simplifying essential work. Selecting efficient tactics that might produce acceptable prediction performance is a challenging task for prediction models. In order to improve classification performance by addressing the misclassification issue, this study proposes a hybrid approach known as arbitrator miniature that combines factor analysis with the following nine machine learning techniques: Support vector machine, Random Forest, K Nearest Neighbor, Logistic Regression, Artificial neural network, Decision Tree, XG boost, Ada boost and Naïve Bayes. To evaluate the robustness of the suggested models, student datasets from a variety of academic fields at diploma-granting institutions in Karnataka, India, were used. In order to assess the proposed model using the datasets, assessment criteria such as classification accuracy and root mean square error were employed. This study’s findings revealed that proposed arbitrator miniature model might significantly improve classification performance. For the purpose of resolving prediction and classification issues, the proposed arbitrator miniature may be viewed as the best prediction models.

Timely categorization as well as prediction of student level of performance provides a heads-up and even a plan for enhancing both the students' and many other administrative settings substandard performance [1]. Therefore, our goal is to address the unidentified student behavior pattern that has an impact on performance. The achievement of students is influenced by a variety of elements, including student background and behavior, assessments, teaching and domain knowledge [2]. These associated elements served as predictive criteria for forecasting student success. In order to track students' academic development, supervised learning was utilized to predict, categorize, and assess their performance. The difficult challenge is selecting the best algorithm that can deliver pleasing outcomes, though. Student success was often reviewed as well as predicted using machine learning methods such as Support vector machine, Random Forest, K Nearest Neighbor, Logistic Regression, Artificial neural network, Decision Tree, XG boost, Ada boost and Naïve Bayes [1–14]. Depending on the dimensions and nature of the data, each model's behavior differs from dataset to dataset. In the classification problem, poor data quality can lead to misclassification, which lowers model accuracy and baffles the algorithms. [1] Released an article in which the author's main objective was to apply principal component analysis to break down the connected observable predictors into a more manageable collection of independent dimensions or components. We used factor analysis as a feature extraction approach to turn the original dataset into a new dataset of high quality data with the intention of delving deeper to investigate and test the model for the existence of hidden variables that may be driving the variance in the data. The models' predictive performance is assessed, and their performance in a fresh dataset, testing samples, or test data is assessed using 10-fold cross-validation. This research intends to propose a machine learning framework called arbitrator miniature for the classification issue. The suggested framework incorporates 10 fold cross-validation, factor analysis, and nine fundamental machine learning algorithms.

Significant objectives of this examination

Applying the arbitrator miniature to the datasets.
To develop arbitrator miniature framework of predication for academic performance of students.
Evaluating arbitrator miniature framework with other ML models & compare their scores like accuracy and RMSE

The efficient prediction model is crucial for supervised machine learning in order to address prediction as well as classification issues. As it was indicated with in Introduction, the domain of educational data mining (EDM) has investigated several machine learning approaches in order to establish how well these approaches can predict students' upcoming academic performance [1] [3] [5]. [1] Used machine learning techniques such support vector machines, decision trees, navies Bayes, and random forests. In order to improve classification performance by addressing the misclassification problem, we expanded it in this work by introducing a hybrid approach of factor analysis in conjunction with nine machine learning algorithms: Support vector machine, Random Forest, K Nearest Neighbor, Logistic Regression, Artificial neural network, Decision Tree, XG boost, Ada boost and Naïve Bayes. The common and cutting-edge classification techniques being used to evaluate student performance are listed in Table I. Finding the most effective algorithms to forecast performance has involved examining a variety of studies.

Although the hold-out approach is more frequently employed in predicting student performance, the k-fold cross-validation is still a robust cross-validation technique. In [23] saw the implementation of and contrasted K-Nearest, Naive Bayes, and Decision Tree while utilizing 10-fold cross-validation and neighbor models. Other studies, including [25] and [24], used k-fold cross-validation to compare various data mining techniques using models to predict student performance.

This is common knowledge how raising a model's forecast accuracy may be difficult. Accurate prediction may be improved by a variety of things. The feature selecting and addressing the misclassification are some of the key elements. Consequently, [26] created the Logistic Regression and Decision Tree models can anticipate students' academic achievement despite managing an uneven class issue. [27] Emphasis was placed on creating diverse utilizing the feature importance technique and algorithms to address the classification issue.

2.1 Identifying research gap

Prognosticating student accomplishment is a challenging endeavor that calls for novel, inventive ways that take into account the changing conditions and factors that affect students' academic performance. The effect of these variables and situations may change from institution to institution and from one class of students to another[51]. Our thorough analysis of the linked studies found deficiencies in the following areas:

Absence of hybrid systems that simplify and enhance the accuracy of predictions of student academic achievement and combine the benefits of supervised and unsupervised learning.

The rigidity of current models when it comes to analyzing various academic and non-academic aspects that are thought to affect how well students learn. While some techniques just take into account a limited portion of relevant components, some approaches forecast student outcomes without linking them to the enabling factors or potential shortcomings.

The validity of the model is questioned since so many student prediction models are verified using a single dataset.

3.1 Datasets

Throughout this research, we made an effort to compile all latent factors influencing students' performance. We used seven datasets to verify the resilience and efficiency of our suggested methods. Details of this dataset are shown in the Table 4. Dataset included one target variable that described the performance levels of students based on their score and 43 characteristics that described the information about each student's learning activities. The features observing from the three primary affected factors make up the predicted features. The forty-four variables made up of these main elements are described in Table II. The target variable's preset classes were listed in Table III. The real dataset, known as AD1, has 1500 samples, which was gathered from 12 polytechnic colleges in Karnataka, India. Use of questionnaires was employed for the data collecting. Students were requested to provide demographic data pertaining to outside influences including family, individual, or student, and school issues. The administrative offices in each school provided the math, it skills and statistics test results for the first semester's students.

3.2 Preprocessing

In this research, preprocessing activities included operations like data cleansing, data transformation, and data discretization. The completion of the questionnaire during data collection included ignoring some items and entering erroneous values (outliers). Since there were few missing values in our datasets, we cleaned the data using the imputing method. We substituted its modes or high frequency category values for the missing value in our categorical variables. We replaced a few missing values and outliers in the output variable with the mean value. We converted several numerical properties into ordinal types for convenience. As indicated in Table III, we discretized the output variables in our study into five performance levels.

3.3 Machine learning models

Machine learning models may be categorize in many different ways. A variety of classifiers are used in this study, including Random Forest [27, 28], K-nearest-neighbor [29, 30], Artificial Neural Network [31, 32], XG-boost [33, 34], Support Vector Machine (Radial Basis Function kernel) [35, 36], Decision Tree [37, 38], Logistic Regression [39, 40], and Nave Bayes [41]. It is commonly known that the majority of machine learning classifiers, including Artificial Neural Networks (ANN), K-Nearest Neighbor (KNN), Random Forests (RF), Logistic Regressions (LR), Decision Trees (DT), and Naive Bayes, naturally allow multiclass classification (NB). Since Support Vector Machine (SVM) and XG-Boost do not naturally support multiclass classification, the Support Vector Machine model is applied using the one-versus-one approach, and the XG-Boost model is applied using the one-versus-all method [26].

Table 4 lists all of the machine learning models that were employed in this study along with the values for each of their unique parameters. Similar parameter setting is utilized in [26]. The parameter setting described in [26] has been adjusted in our study to better fit our issue description. To create predictions, these factors are essential.

Besides incorporating factor analysis as well as machine learning technologies, arbitrator model is presented. At initial stage baseline models are suggested. With k-fold cross-validation, performance of suggested baseline models are enhanced. Finally, by merging factor analysis with machine learning, as shown in Fig. 1, arbitrator model is suggested.

4.1 Baseline Models

A number of efficient machine learning techniques which have been widely used in academic settings. We need to use a variety of machine learning approaches in learning environments, including association rule mining, regression analysis, classification, and clustering [3]. In classifying and predict the target variables' preset classes, classification is a typical machine learning approach. In this study, we looked at a number of machine learning classifiers and chose five cutting-edge techniques that are often used to forecast academic success [3–14]. This study proposes a hybrid approach known as arbitrator miniature that combines factor analysis with the following five machine learning techniques: Support Vector Machine, Adaboost, Gradient Boost, Random Forest, and Logistic Regression.

4.2 Cross Validation

A quantitative approach used to gauge how well machine learning algorithms work is cross-validation. There are different cross-validation techniques, but the k-fold cross validation is used because it is well-liked and simple to comprehend, and because it often produces less distortion than the other techniques. 10-fold cross-validation approach is used in this study, also known as 10-CV, to reach proposed model.

4.3 Arbitrator Miniature

In supervised machine learning, classification comprises the preponderance of tasks. A challenging topic in data mining and machine learning is the classification problem. We suggested the four renowned classifiers with lots of advantages. Nevertheless, the main the issue with all of above mentioned classifiers is noisy data and over fitting, which makes it possible to determine the correctness of the classification. In an effort to solve this issue, we strive to unnecessary and unrelated traits that are disruptive in the process of categorization. When analyzing data, it calls for more while using computing resources, it takes a lot of time. Data does have a big volume. Because of this, the feature extraction method for removing noise from data to save time and reduce resource use and restore high-quality data. By integrating a dimensional reduction with classification approaches, the accuracy and efficiency might be enhanced.

The process of converting the initial dataset X of size l, which contains potentially related features, into a new dataset Z of size m(m < l), which contains linear unrelated characteristics are as follows:

Step 1: Find the mean of each feature based on the previously processed data using the following equation:

\(\mu =1/n{\sum }_{i=0}^{n}*\) x _i (1)

Step 2: Calculate the variance using below formula in order to explore and deviation of each feature in the dataset.

Var(X) = σ _x ² = \(1/n{\sum }_{i=0}^{n}*(\)x_i-\(\mu\)) ² (2)

Step 3: Eq. (3) is used to determine the covariance and correlation for two variables, labelled X and Y.

Cov(X,Y) = σ_x² = \(1/n{\sum }_{i=0}^{n}*(\)x_i-µ_x ) \((\)y_i-µ_y) (3)

Step 4: The mean of the eigenvectors and eigenvalues used to describe the characteristics in the multiple datasets. The acquired eigenvectors will indicate the whereas the eigenvalues represent its magnitude, new features space. It is feasible to find the eigenvalues by solving the equation.

Det(S-λI) = 0 (4)

We need to select a dataset with a range of feature sets that is related to the students' academic performance. Any and all inconsistencies that could have existed in the dataset at the time the data was gathered must be removed during the preparation stage. It is now time to test our dataset in two different ways: the first mode will test our dataset using all of the characteristics that are now there, and the second mode will pick the factors to test using proposed arbitrator miniature. In this case, we find the accuracy of the machine learning algorithm in three phases .Phase 1 is to find the accuracy of baseline models. Phase 2 is to improve the accuracy of baseline models by 10-fold cross-validation (10- CV). Phase 3 is to execute proposed arbitrator miniature which is the combination of the baseline models with 10-CV and FA.

The assessment of the pictorial confusion matrix can be used to assess how well each proposed model performed in analyzing and forecasting student performance. Our output variable can be divided into five ordinal categories, as shown in Table I, without losing coherence. The proportion of successfully predicted outcomes is measured by accuracy. By evaluating the percentage of accurately estimated student performance levels, we can assess the potential of our prediction model in this case [15].

Accuracy = \(\frac{\sum \varvec{X}}{\sum \varvec{Y}}\)* 100 (5) RMSE = \(\sqrt{{\sum }_{i=1}^{m}\frac{\left(X- Y\right)2}{\text{M}}}\) (6)

Our objective is to determine how near our forecast are to the academic achievement of the students as well as their skill levels. These ordinal skill sets, average, good, and excellent—were encoded as 1, 2, 3, and 4, respectively. Calculating the RMSE involves: The predicted skill level is given by Y= {1, 2, 3,4} and the actual performance level is given by X={1, 2,3,4}. In contrast to accuracy, the model is better the smaller the RMSE. The prediction model is perfect if the RMSE is equal to 0.

7.1 Experimental setup

The goal of this research is to construct a unique ensemble learning-based intelligent prediction model to forecast student performance. The suggested strategy made use of cross validation, machine learning, and baseline model combinations. The classification results of these classifiers were subjected to ensemble learning. In order to evaluate the effectiveness of ensemble learning-based classifiers, tests were carried out on the UCI repository using a Python program with Tenser Flow, Keras, and other related libraries. On an 11th generation Intel Core i7 with two NVIDIA GeForce RTX 3060 Laptop GPUs with 6.0 GB and 7.9 GB of RAM, research was conducted.

7.1 Sources of datasets

We used seven datasets to verify the resilience and efficiency of our suggested methods. Details of this dataset are shown in the Table 4, first four datasets are used to access the whole approach and our focus is on predicting student's grades and factors for only the next upcoming semester.

The remaining datasets—Emotions, Flags, and Stanford background—used in our investigations are related to non-academic fields including music, audio, and photos but serve as a standard for measuring how well our arbitrator miniature performs. As previously indicated, it might be challenging to locate the necessary scholarly datasets for this strategy in readily accessible sources. Additionally, none of the associated university-level research that we are aware of provided their datasets for experimental replication, most likely due to data privacy limitations [51].

The chosen non-academic dataset’s statistical properties may be found in [52][53][54]. According to our defined dataset, shown in Fig. 2, these dataset kinds are valid. These benchmarked datasets comprise multi-label numerical outputs that are specifically abstract representations of various rates of influential elements. As a result, they may be applied to generally evaluate performance. Additionally, these datasets are acceptable because they offer a number of crucial qualities, including data density, cardinality, and distinct [50][51].

The three datasets, DS1, DS2 and DS3 are produced datasets that were built using suggested characteristics for the output variable as indicated in [18–20]. Fourth one is the real dataset, known as AD1, which was gathered from 12 polytechnic colleges in Karnataka, India. Use of questionnaires was employed for the data collecting. Investigations are conducted on three different stages. Stage 1 involves putting the results of the baseline models into practice. Stage 2 involves a 10-fold cross-validation improvement of the baseline models (10- CV). Stage 3 involves putting the proposed model—a fusion of the baseline models with 10-CV and FA—into action.

Table 4

Description of datasets used in the study
Datasets	Source	Type	Samples	Features	Labels
DS1	Generated	Academic	2000	30 (Domestic Factors)	5
DS2	Generated	Academic	1000	5 (Soft Skill factors)	5
DS3	Generated	Academic	3000	30 (Individual factors and School factors)	5
AD1	Real	Academic	1500	Combination of important factors from DS1, DS2 and DS3	5
Emotions	Kaggle	Non Academic	593	72	6
Flags	Kaggle	Non Academic	194	10	7
Stanford Background	Kaggle	Non Academic	2407	294	6

7.2 Implications of Baseline Models

We put forward the nine most well-liked machine learning methods:.. The tables display the two performance measures, classification accuracy and RMSE. Table 5 revealed that the logistic regression model was the least, but the Random Forest approach produced the best results in terms of classification accuracy and RMSE, indicating that it might be a good model. : Support vector machine, Random Forest, K Nearest Neighbor, Logistic Regression, Artificial neural network, Decision Tree, XG boost, Ada boost and Naïve Bayes.

Table 5

Comparison of different classification Algorithms with their Prediction Accuracy for DS1 dataset
ML Algorithms	Accuracy (Baseline models)	Accuracy (Baseline models + 10-cv )	Accuracy (Proposed Model)
Support Vector Machine	85.36%	86.66%	87.32%
Ada boost	73.50%	75.34%	80.50%
XG Boost	74.45%	75.48%	81.33%
Random Forest	85.71%	87.28%	91.79%
Logistic Regression.	60.76%	72.66%	80.57%
Artificial Neural Network	77.64%	83.62%	87.75%
K Nearest Neighbor	75.87%	81.43%	88.65%
Decision Tree	80.65%	85.34%	90.45%
Naïve Bayes	65.76%	77.45%	83.65%

Table 6

Comparison of different classification Algorithms with their Prediction Accuracy for DS2 dataset
Ml Algorithms	Accuracy (Baseline models)	Accuracy (Baseline models + 10-cv )	Accuracy (Proposed Model)
Support Vector Machine	85.96%	87.76%	88.46%
Ada boost	74.10%	77.94%	81.98%
XG Boost	74.89%	76.41%	83.67%
Random Forest	86.11%	89.88%	92.70%
Logistic Regression.	61.76%	73.09%	83.07%
Artificial Neural Network	76.68%	82.12%	88.75%
K Nearest Neighbor	77.34%	83.33%	87.65%
Decision Tree	83.62%	86.56%	87.45%
Naïve Bayes	68.16%	79.56%	83.65%

Table 7

Comparison of different classification Algorithms with their Prediction Accuracy for DS3 dataset
ML Algorithms	Accuracy (Baseline models)	Accuracy (Baseline models + 10-cv )	Accuracy (Proposed Model)
Support Vector Machine	86.16%	88.21%	91.12%
Ada boost	74.94%	79.42%	85.67%
XG Boost	76.29%	80.81%	87.45%
Random Forest	87.15%	90.63%	95.89%
Logistic Regression	63.36%	79.09%	82.33%
Artificial Neural Network	78.78%	87.10%	90.89%
K Nearest Neighbor	79.04%	85.56%	89.09%
Decision Tree	82.60%	89.87%	93.34%
Naïve Bayes	70.16%	77.58%	84.31%

7.2 Implications of Baseline Models with k-fold Cross-Validation

In prediction and classification models, the k-fold cross-validation approach is frequently used to divide the dataset into k-1 sub folds for training sets and 1 fold for testing sets, then rotate the folds. Since it functions best at this split, we performed 10-fold cross validation in this trial. 10% of the data was utilized for testing, while 90% was used in the training stage. The average of all assessment criteria is then calculated when all interactions have been completed. The accuracy of SVM was increased by 4% as shown in Table 4. The weak logistic regression classifier’s performance was subsequently markedly enhanced to 80.66%.

Table 8

Comparison of different classification Algorithms with their RMSE for DS1 dataset
ML Algorithms	RMSE (Baseline models)	RMSE (Baseline models + 10-cv )	RMSE (Proposed Model)
Support Vector Machine	0.701	0.691	0.410
Ada boost	1.033	0.914	0.681
XG Boost	1.134	0.721	0.521
Random Forest	0.602	0.474	0.321
Logistic Regression.	1.164	0.931	0.664
Artificial Neural Network	0.876	0.603	0.489
K Nearest Neighbor	0.908	0.827	0.532
Decision Tree	0.779	0.546	0.369
Naïve Bayes	1.023	0.943	0.787

Table 9

Comparison of different classification Algorithms with their RMSE for DS2 dataset
ML Algorithms	RMSE (Baseline models)	RMSE (Baseline models + 10-cv )	RMSE (Proposed Model)
Support Vector Machine	0.611	0.521	0.401
Ada boost	0.903	0.804	0.654
XG Boost	1.022	0.771	0.489
Random Forest	0.634	0.454	0.309
Logistic Regression.	1.055	0.841	0.598
Artificial Neural Network	0.89	0.765	0.633
K Nearest Neighbor	0.9	0.667	0.609
Decision Tree	0.8	0.698	0.577
Naïve Bayes	0.955	0.9	0.799

Table 10

Comparison of different classification Algorithms with their RMSE for DS3 dataset
ML Algorithms	RMSE (Baseline models)	RMSE (Baseline models + 10-cv )	RMSE (Proposed Model)
Support Vector Machine	0.7	0.554	0.266
Ada boost	0.89	0.676	0.356
XG Boost	0.877	0.6	0.334
Random Forest	0.614	0.289	0.119
Logistic Regression.	0.767	0.556	0.455
Artificial Neural Network	0.745	0.599	0.476
K Nearest Neighbor	0.776	0.567	0.324
Decision Tree	0.696	0.435	0.235
Naïve Bayes	0.8	0.676	0.514

7.3 Implications of Proposed mode1

By combining the baseline models with a feature reduction strategy called FA, we were able to create the fusion models we have suggested. One of the effective techniques in classification models for eliminating unrelated or unnecessary features is feature extraction. Dimensionality reduction by FA [13][21] may most certainly be used as regularization to avoid over fitting and boost model accuracy. People frequently fall into the trap of believing that FA chooses certain characteristics from the dataset while discarding others. Actually, the algorithm creates a fresh dataset of attributes by combining the previous ones. Table 4 demonstrates how the suggested model helps classifiers become more accurate.

Table 11

Comparison of different classification Algorithms with their Prediction Accuracy for AD1 dataset
ML Algorithms	Accuracy (Baseline models)	Accuracy (Baseline models + 10-cv )	Accuracy (Proposed Model)
Support Vector Machine	87.46%	91.66%	93.32%
Ada boost	75.57%	83.34%	89.50%
XG Boost	77.03%	84.48%	91.33%
Random Forest	89.73%	93.28%	97.79%
Logistic Regression.	67.56%	80.66%	86.57%
Artificial Neural Network	83.78%	88.90%	93.89%
K Nearest Neighbor	79.97%	83.14%	92.78%
Decision Tree	84.09%	91.97%	95.07%
Naïve Bayes	73.45%	81.59%	90.31%

Figures 3, 4, 5, 9 showed how each model performed according to its accuracy during each step. We discovered that the 10-CV improvement in conjunction with PCA produces the greatest results in forecasting student performance. The Figs. 6, 7, 8, 10 display how well the models' RMSE performed at each phase. The proposed models that have been suggested could produce relatively little RMSE. In this prediction scenario, the hybrid RF algorithm provided the least value of RMSE, demonstrating its superiority as the best predictive model. According to the findings, we may enhance the performance of our basic models by employing 10-CV. We also noticed that the innovative fusion models that were suggested may improve classification performance and produce better outcomes. The suggested proposed models can be viewed as the best prediction models for resolving categorization and prediction issues.

7.4 Analysis of the proposed Approach in relation to the existing Approaches

Multiple classifiers have been used in numerous research papers in EDM to predict student achievement.[55] recently put up an ensemble model to recognize at-risk kids and give them guidance on managing their learning. They combined four ensemble algorithms—bagging, random subspace, multilayer perceptron, and random forest—with four single classifiers. The evaluation's findings revealed that the ensemble model had an accuracy rate of 91.70%. Another study used logging data to identify at-risk children by estimating their learning success based on their learning habits. Along with Random Forest, Multilayer Perceptron, and Gaussian Naive Bayes, they employed Logistic Regression. The outcomes demonstrated that Random Forest outperformed other models including the baseline Logistic Regression with 89% accuracy [56]. A model to forecast pupils' achievement based on their daily actions was introduced by [57]. Using data mining techniques, an AA model has been proposed to assess institutional performance based on key performance metrics. The results revealed that, in compared to other machine learning models used in the study, artificial neural networks performed better in terms of accuracy (82.9%) [59]. To improve the effectiveness of classifiers, ensemble approaches including bagging, boosting, and random forests were used to predict student performance in a learning management system based on behavioral variables. To improve academic performance, ensemble techniques were used to the classifiers, yielding an accuracy of 91.5% [60]. [61] Proposed a forecast model for students' performance based on data mining. A decision tree, logistic regression, naive Bayes tree, artificial neural network, support vector machine, and k-nearest neighbor are a few of the data mining techniques used to assess student performance. These classifiers employed ensemble approaches including bagging, boosting, random forest, and voting to increase their output. According to the findings, bagging enhanced the decision tree algorithm's accuracy from 90.4–91.4%. Similar improvements were made to RMSE findings, which went from 0.904 to 0.914, and to precision results, which went from 0.905 to 0.914. The suggested model includes four well-known ensemble approaches, including bagging, boosting, stacking, and voting, in addition to nine conventional machine learning algorithms. The NB model assessed the RMSE score at 0.71% and 0.75%, respectively, by integrating boosting and GBT with AdaBoost [61]. [62] Research centered on the use of data mining tools and ensemble approaches to predict students' performance. Additionally, they put forth brand-new hybrid classifiers to produce precise forecasts of student performance. In contrast to basic classifiers and ensemble approaches used in the same research, the findings indicated that the hybrid model surpassed the other classifiers in terms of accuracy (i.e., 81.67). In contrast to cutting-edge ensemble methodologies suggested in EDM, the fusion ensemble-based strategy used in this research study to boost academic performance achieved the greatest accuracy (i.e., 97.79%) and RMSE 0.119. As a consequence, the study's findings show that the suggested prediction model is reliable. In comparison to previous approaches that stress how the fusion of ensemble techniques might increase the proportion of prediction, our approach performs better across the board [62].

7.5 Threats to Validity and Research Limitations

We highlight the key risks that might possibly impair the internal and external validity of our method with regard to the viability of employing the suggested hybrid/multi-label classier models and the validity of the conducted experiments. Threats to internal validity may come from biases used in designing the studies. However, using non-real datasets for studies poses a possible danger to external validity [51]. By reporting on the results that are an average of 10 similarly designed runs, we have lessened the potential of unexpected biases that may be created when we configured our trials. We additionally handle this issue by configuring the Random Number Generator in Python since initializing weight coefficients randomly may lead to different measurements for each run.

Regarding the universality of our strategy, locating relevant and extensive datasets that encompass student outcomes over many years and for many degrees presents a substantial problem when attempting predictive modelling in the area of teaching and learning. We utilized the greatest publicly available free datasets we could locate, which adds sampling bias. It goes without saying that such small datasets do not account for all potential variables that can affect student progress. For instance, despite its indisputable significance, student participation is neither available nor taken into consideration. The databases show how students with certain majors performed, which may be different from those pursuing other degrees, like psychology. Additionally, we verified our model using non-academic datasets, which can show patterns different from those seen in actual educational datasets [51].

Academic achievement for students is an inter phenomenon that may be analyzed from several perspectives. It is uncertain if the model will be capable of predicting other measures of academic success, such as scores on standardized tests and the percentage of students who achieve their goals, with the same level of accuracy, particularly when several measures are combined to determine a student's success [51]. Undoubtedly, this is a problem for future investigation. Furthermore, we haven't conducted any additional testing using alternative datasets to verify the model's dependability.

As a result, we were unable to assert the universality of our technique. However, it could be essential to adjust the hybrid/multi-label classifier models to account for different educational contexts and datasets. Future research will use authentic academic datasets to investigate the highlighted constraints in further depth.

The nine well-known machine learning classifiers for predicting student performance were introduced in this study. The nine suggested algorithms are:. Support vector machine, Random Forest, K Nearest Neighbor, Logistic Regression, Artificial neural network, Decision Tree, XG boost, Ada boost and Naïve Bayes. Three stages make up the operation. First, we evaluated how well various baseline techniques performed. Second, we used 10-CV to boost performance. Finally, to enhance classification performance, we integrated the 10-CV approach with FA to baseline models. Based on classification accuracy and RMSE as measurement parameters, it can be seen that the suggested arbitrator miniature models gave highly pleasing results when combined with FA and 10-CV. In conclusion, the proposed models provided a high performance that demonstrates itself as a possible approach for solving the problem by merging the baseline models with factor analysis and assessed by k-fold cross-validation. We consider a thorough examination of algorithms using different performance indicators since it helps us comprehend the algorithms better. Different performance criteria, including accuracy and RMSE, were taken into account. By contrasting several algorithms using various criteria, it is possible to assess an algorithm's predictability. The literature study indicates that because each problem has specific features, it is challenging to forecast which performance metrics are better for certain challenges. As a result, combining numerous measurements is advised for improved algorithm performance. Using several feature selection techniques, we examined the effectiveness of various selected data mining methods on the relevant dataset. Finally, we can state that any classification algorithm's performance when combined with feature prediction gives better result. This study's methodology will help educational leaders and legislators create new regulations and curricula pertaining to student retention in higher education. Additionally, by identifying students who are at danger of quitting school early, this research enables prompt assistance and intervention. We compare the performance accuracy and efficiency of individual classifiers and classifier ensembles. This model can assist academics in identifying children with learning difficulties, developing the learning experience, and lowering educational failure rates. It can also assist managers in managing more effectively depending on the outcomes of the learning system. The model could be expanded, refined in the future to encompass a broad range of student dataset attributes.

Declarations

Conflicts of interest/Competing interests:

The authors declare that there are no conflicts of interest or competing interests.

Funding:

Not applicable.

Herbert K., "The New Book of Popular Science", World Applied Sciences Journal, Daribury, Connecticut: Grolier Inc., 1978.
Ministry of Education, Youth and Sport, “Education in Cambodia: Finding from Cambodia's Experience in PISA for Development”, Phnom Penh: Author, 2018.
S. Slater, S. Joksimovic, V. Kovanovic, R.s Baker, and D. Gasevic, "Tools for Educational Data Mining: A Review", Journal of Educational and Behavioral Statistics, Vol. 42, No. 1, 2016, pp. 88–106.
Pooja Thakar, Anil Mehta, and Manisha, "Performance Analysis and Prediction in Educational Data Mining: A Research Travelogue", International Journal of Computer Application, Vol. 100, No.12, January 2015, pp. 60–68.
C. Romero and Ventura., “Educational Data Mining: A Review of the State of Art”, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, vol. 40, No. 6, 2010, pp. 601–618.
Akinrotimi A.O, and Aremu D.R, "Student Performance Prediction Using Randomtree and C4.5," Journal of Digital Innovation and Contemporary Research, Engineering and Technology, Vol. 6, No. 3, 2018, pp. 23–34.
Amjad A. S., Mostafa Al-Emran, and Khaled S., “Mining Student Information System Records to Predict Students’ Academic Performance”, Springer Nature Switzerland AG 2020, AMLTA 2019, AISC 921, 2019, pp. 229–239.
Kotsiantis S., Piarrekeas C., and Pintelas P., "Predicting Students' performance in Distance Learning using Machine Learning Techniques", Applied Artificial Intelligence, Vol. 18, 2007, pp. 411–426.
Hamoud A. K., Hashim A. S., and Awadh W. A., "Predicting Student Performance in Higher Education Institutes Using Decision Tree Analysis", International Journal of Interactive Multimedia and Artificial Intelligent, Vol. 5, No. 2, February 2018, pp. 26–31.
Daud A., Aljonhani N.R., Abbasi R.A., Lytras M.D., Abbas F., Alowibdi J.S., “Prediction student performance using advanced learning analytics”, Proceedings of 26th International Conference on World Wide Web, Companion, Perth, Australia, April 2016, pp. 416–421.
M.S. Mythili,, A.R.M. Shanavas, “An Analysis of Students' Performance using Classification Algorithms”, IOSR Journal of computer Engineering (IOSR-JCE), Vol 16, No. 1, January 2014, pp. 63–69.
Aung Nway Oo, “Comparative Study of Principle Component Analysis based on Decision Tree Algorithm”, International Journal of Advances in Scientiofic Research and Engineering, Vol. 4, No. 6, June 2018, pp. 122–126.
Karthikeyan T., Thangaraju P., “PCA-NB Algorithms to Enhance the Predictive Accuracy”, International Journal of Engineering and Technology, Vol. 6, No. 1, 2014, pp. 381–387.
Farid J., Ahmad A.S., “Building student’s performance cesissiion tree classifier using boosting algorithm”, Indonesian Journal of Electrical Engineering and Computer Science, Vol. 14, No. 3,2019, pp. 1298–1304.
Babak M.A., Seyed K.S., Maryam M.M., “Support vector machinebased arrhythmia classification using reduced features of heart rate variablitiy singanal”, Ariticial Intelligence in Mechine (Elsevier), vol. 44, 2008, pp. 51–64.
Humera S., Raniah Z., Kavitha G., “Prediction of Student Performance in Semester Exam Using a Naïve Bayes Classifier”, International Journal of Innovative Research in Science, Engineering and Technology, Vol. 4, No. 10, October 201 5, pp. 9823–9829.
Jollifie I.T, "Principle components analysis and factor analysis", Springer 1986.
Phauk Sokkhey and Takeo Okazaki., "Comparative Study of Prediction Models on High School Student Performance in Mathematics", Journal of IEIE Transaction on Smart Processing and Computing, Vol. 8, No. 5, October 2019, pp. 394–404.
Mohamed Z.G. A., Mustafa B. M., Lazim A., and Hamdan A. M., "The Factors Influence Students' Achievement in Mathematics: A Case for Libyan's Students ". Australian Journal of Basic and Applied Science, Vol. 17, N0. 9, 2012, pp. 1224–1230.
Uysal S., "Factors affecting the Mathematics achievement of Turkish students in PISA 2012", Academic Journals, Vol. 10, June 2015, pp. 1670–1678
Phauk Sokkhey and Takeo Okazaki., "Hybrid Machine Learning Algorithms for Predicting Academic Performance”, International Journal of Advanced Computer Science and Applications (IJACSA), Volume 11 Issue 1, 2020.
Bindhia K. Francis, Suvanam Sasidhar Babu, “Predicting Academic Performance of Students Using a Hybrid Data Mining Approach”, Journal of Medical Systems, Volume 43 Issue 6 Jun 2019.
Mukesh Kumar, Nidhi, Bhisham Sharma, Disha Handa, "Building Predictive Model by Using Data Mining and Feature Selection Techniques on Academic Dataset", International Journal of Modern Education and Computer Science(IJMECS), Vol.14, No.4, pp. 16–29, 2022.DOI: 10.5815/ijmecs.2022.04.02.
Bridgitte Owusu-Boadu1, Isaac Kofi Nti2*, Owusu Nyarko-Boateng3, Justice Aning4, Victoria Boafo5,” Academic Performance Modelling with Machine Learning Based on Cognitive and Non-Cognitive Features”, Applied Computer Systems ISSN 2255–8691 (online) ISSN 2255–8683 (print) December 2021, vol. 26, no. 2, pp. 122–131.
Polyzou, A., & Karypis, G. (2016). Grade prediction with models specific to students and courses. International Journal of Data Science and Analytics, 2(3–4), 159–171.
R. Ghorbani, R. Ghousi,,” Comparing Different Resampling Methods inPredicting Students' Performance Using Machine Learning Techniques”, April 2020, IEEE Access 8(1):67899–67911.
A. K. Jain, J. Mao, and K. M. Mohiuddin, ``Arti_cial neural networks: A tutorial,'' Computer, vol. 29, no. 3, pp. 31_44, Mar. 1996.
G. Zhang, B. E. Patuwo, and M. Y. Hu, ``Forecasting with arti_cial neural networks:: The state of the art,'' Int. J. Forecasting, vol. 14, pp. 35_62,Mar. 1998.
P. Cunningham and S. J. Delany, ``k-Nearest neighbour classi_ers,'' Mul-tiple Classi_er Syst., vol. 34, pp. 1_17, Mar. 2007.
J.Wang, P. Neskovic, and L. N. Cooper, ``Improving nearest neighbor ruwith a simple adaptive distance measure,'' Pattern Recognit. Lett., vol. 28, no. 2, pp. 207_213, 2007.
A. K. Jain, J. Mao, and K. M. Mohiuddin, ``Arti_cial neural networks: A tutorial,'' Computer, vol. 29, no. 3, pp. 31_44, Mar. 1996.
I. A. Basheer and M. Hajmeer, ``Arti_cial neural networks: Fundamentals, computing, design, and application,'' J. Microbiol. Methods, vol. 43, no. 1, pp. 3_31, Dec. 2000.
M. Zi¦ba, S. K. Tomczak, and J. M. Tomczak, ``Ensemble boosted trees with synthetic features generation in application to bankruptcy prediction,'' Expert Syst. Appl., vol. 58, pp. 93_101, Oct. 2016.
Y. Freund, R. Schapire, and N. Abe, ``A short introduction to boosting,'' J.-Jpn. Soc. Artif. Intell., vol. 14, nos. 771_780, p. 1612, 1999.
C.-W. Hsu, C.-C. Chang, and C.-J. Lin, ``Apractical guide to support vector classi_cation,'' Dept. Comput. Sci. Inf. Eng., Univ. Nat. Taiwan, Taipei,Taiwan, Tech. Rep., 2003, pp. 1_12.
J. A. Suykens and J. Vandewalle, ``Least squares support vector machine classi_ers,'' Neural Process. Lett., vol. 9, no. 3, pp. 293_300, 1999.
S. R. Safavian and D. Landgrebe, ``A survey of decision tree classifier methodology,'' IEEE Trans. Syst., Man, Cybern., vol. 21, no. 3,pp. 660_674, 1991.
W. Du and Z. Zhan, ``Building decision tree classier on private data,'' in Proc. IEEE Int. Conf. Privacy, Secur. Data Mining, vol. 14. Darlinghurst, NSW, Australia: Australian Computer Society, 2002, pp. 1_8.
D. R. Cox, ``The regression analysis of binary sequences,'' J. Roy. Stat. Soc., B, Methodol., vol. 20, no. 2, pp. 215_232, Jul. 1958.
D. W. Hosmer, Jr., S. Lemeshow, and R. X. Sturdivant, Applied Logistic Regression. Hoboken, NJ, USA: Wiley, 2013.
I. Rish, ``An empirical study of the naive Bayes classier,'' in Proc. Workshop Empirical Methods Artif. Intell., vol. 3, no. 22, pp. 41_46, 2001.
A. Polyzou and G. Karypis, ``Grade prediction with models specific to students and courses,'' Int. J. Data Sci. Analytics, vol. 2, nos. 3_4, pp. 159_171, Dec. 2016.
S. J. H. Yang, O. H. T. Lu, A. Y. Q. Huang, J. C. H. Huang, H. Ogata, and A. J. Q. Lin, ``Predicting Students' academic performance using multiple linear regression and principal component analysis,'' J. Inf. Process., vol. 26, pp. 170_176, 2018.
L. M. A. Zohair, ``Prediction of student's performance by modelling small dataset size,'' Int. J. Educ. Technol. Higher Edu., vol. 16, no. 1, p. 27, 2019.
Q. Hu and H. Rangwala, ``Reliable deep grade prediction with uncertainty estimation,'' in Proc. 9th Int. Conf. Learn. Anal. Knowl., 2019, pp. 76_85.
Q. Hu and H. Rangwala, ``Academic performance estimation with attention-based graph convolutional networks,'' 2019, arXiv:2001.00632.[Online]. Available: http://arxiv.org/abs/2001.00632.
A. Daud, N. R. Aljohani, R. A. Abbasi, M. D. Lytras, F. Abbas, and J. S. Alowibdi, ``Predicting student performance using advanced learning analytics,'' in Proc. 26th Int. Conf. World Wide Web Companion, 2017,pp. 415_421.
F. Marbouti, H. A. Diefes-Dux, and K. Madhavan, ``Models for early prediction of at-risk students in a course using standards-based grading,''Comput. Edu., vol. 103, pp. 1_15, Dec. 2016.
G. G. Colombini, I. B. M. de Abreu, and R. Cerri, ``A self-organizing map-based method for multi-label classi_cation,'' in Proc. Int. Joint Conf. Neural Netw. (IJCNN), May 2017, pp. 4291_4298.
N. Saini, S. Saha, and P. Bhattacharyya, ``Incorporation of neighborhood concept in enhancing SOM based multi-label classi_cation,'' in Proc. Int. Conf. Pattern Recognit. Mach. Intell. Cham, Switzerland: Springer, 2019, pp. 91_99.
A. Alshanqiti and A. Namoun, "Predicting Student Performance and Its Influential Factors Using Hybrid Regression and Multi-Label Classification," in IEEE Access, vol. 8, pp. 203827–203844, 2020, doi: 10.1109/ACCESS.2020.3036572.
https://.
https://www.kaggle.com/datasets/edoardoba/world-flags.
https://.
Hassan, H.; Ahmad, N.B.; Anuar, S. Improved students’ performance prediction for multi-class imbalanced problems using hybrid and ensemble approach in educational data mining. J. Phys. Conf. Ser. 2020, 1529, 052041.
Wasif, M.; Waheed, H.; Aljohani, N.R.; Hassan, S.-U. Understanding student learning behavior and predicting their performance. In Cognitive Computing in Technology-Enhanced Learning; IGI Global: Hershey, PN, USA, 2019; pp. 1–28.
Emmanuel, A.A.; Aderoju, M.A.; Falade, A.A.F.; Atanda, A. An appraisal of online gambling on undergraduate students’ academic performance in university of Ilorin, Nigeria. Int. J. Innov. Technol. Integr. Educ. 2019, 3, 45–54.
Yousafzai, B.K.; Hayat, M.; Afzal, S. Application of machine learning and data mining in predicting the performance of intermediate and secondary education level student. Educ. Inf. Technol. 2020, 25, 4677–4697.
Alam, T.M.; Mushtaq, M.; Shaukat, K.; Hameed, I.A.; Sarwar, M.U.; Luo, S. A Novel Method for Performance Measurement of Public Educational Institutions Using Machine Learning Models. Appl. Sci. 2021, 11, 9296.
Ragab, M.; Aal, A.M.K.A.; Jifri, A.O.; Omran, N.F. Enhancement of Predicting Students Performance Model Using Ensemble Approaches and Educational Data Mining Techniques. Wirel. Commun. Mob. Comput. 2021, 2021, 6241676.
Saleem, F.; Ullah, Z.; Fakieh, B.; Kateb, F. Intelligent Decision Support System for Predicting Student’s E-Learning Performance Using Ensemble Machine Learning. Mathematics 2021, 9, 2078.
Siddique, Ansar, Jan, Asiya, Majeed, Fiaz, Qahmash, Adel, Quadri, Noorulhasan Naveed, Wahab, Mohammad, 2021/12/13, SP – 11845 “Predicting Academic Performance Using an Efficient Model Based on Fusion of Classifiers” 11, DO – 10.3390/app112411845 Applied Sciences

Author Contribution Statement: Prof. Saleem malik and Dr. Jothimani wrote the main manuscript text and both the authors reviewed the manuscript.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Arbitrator Miniature: A Paradigm using Data Science Methods to Predict Academic Performance

Status:

Version 1

Abstract

1 Introduction

2 Related Works

2.1 Identifying research gap

3 Materials and Methods

3.1 Datasets

3.2 Preprocessing

3.3 Machine learning models

4 Proposed Arbitrator Model

4.1 Baseline Models

4.2 Cross Validation

4.3 Arbitrator Miniature

5 Evaluation Metrics

7 Result Analysis

7.1 Experimental setup

7.1 Sources of datasets

7.2 Implications of Baseline Models

7.2 Implications of Baseline Models with k-fold Cross-Validation

7.3 Implications of Proposed mode1

7.4 Analysis of the proposed Approach in relation to the existing Approaches

7.5 Threats to Validity and Research Limitations

8 Final Thoughts and Way Forward

Declarations

Declarations

Declarations

Funding:

References

Unsectioned Paragraphs

Additional Declarations

Status:

Version 1