The literature offers various mathematical applications, wherein the Naïve Bayes algorithm stands out due to its large application percentage. In this section, we will discuss the success percentages of each study that used the Naïve Bayes algorithm as well as their characteristics, such as the quantity of information required and under what circumstances each study was conducted.

In their research, Kumar et al. (2020) assessed student performance in technical exams used for university application, an important process for universities. Forecasting whether students would pass or fail a technical exam based on their performance in certain subjects of their course was the technique used to approach the problem. The authors used data regarding 200 K L University students from 2013 to 2017. Students were sorted and classified using two different algorithms: K-means clustering and hierarchical clustering. Once classified, the authors used the Naïve Bayes algorithm for their predictions at 72% accuracy.

In their study, Jayaprakash et al. (2020) discuss student dropout rates in the early stages of the program by calculating the factors that exert the most influence on academic student performance. For these predictions, the authors used 887 instances and 19 variables. Additionally, they used the Naïve Bayes algorithm at 85.7% accuracy.

In their work, Daud et al. (2017) assessed the reasons why many students drop out of their careers. For this, the study predicted personal, family, and economic factors that most influence student performance to determine whether any given student would complete their studies. For testing, the dataset is cleaned to obtain 50 students who finished their studies and 50 who did not. As part of their results, the Naïve Bayes algorithm achieved 84.8% in the F1-score.

The study reported by Widyaningsih et al. (2019) intended to forecast academic performance for freshman students using a semisupervised learning approach to classify student performance. In this study, the Naïve Bayes classifier was used as a methodology, which achieved 96% accuracy.

Devasia et al. (2016) addressed student desertion in higher education. The proposed system is a web-based application that employs the Naïve Bayesian mining technique to extract useful information. According to this study, the basic algorithm provides more accuracy in this field than other methods such as regression, decision tree, and neural networks.

Amazona and Hernandez (2019) forecast student performance to improve decision-making education. Educational data mining was used to model student academic performance through the Naïve Bayes algorithm, assess the dataset, and predict academic performance. Results reveal that even when the Naïve Bayes classifier is outperformed by other algorithms, it achieved an overall prediction rate of 88% in classification accuracy.

In their research, Páez and Guzmán (2018) solved problems affecting several teaching and learning processes. To this end, they designed and automated a predictive model of student academic performance. The study was based on several algorithms, wherein the Naïve Bayes outperformed the others at a 73% prediction rate.

In some cases, clustering is required for further data analysis. The aforementioned studies are aware of this fact because K-means is used in Kumar et al. (2020) and Widyaningsih et al. (2019), hierarchical in Kumar et al. (2020), Daud et al. (2017), and Widyaningsih et al. (2019), and custom clustering types in Páez and Guzmán (2018) and others depending on the desired level of classification. Nevertheless, the most common and easiest to use is Naïve Bayes, which was used in Jayaprakash et al. (2020), Widyaningsih et al. (2019), and Amazona and Hernandez (2019). The type of data available is critical because more specific data will better cover student learning, academic environment issues, and more. Thus, as mentioned in Widyaningsih et al. (2019), groups must be as differentiated as possible.

Kumar et al. (2020) also proposed using the C5.0 algorithm, a decision tree algorithm, to predict student performance. For this, they used the exact same methodology as with Naïve Bayes. As a result, they obtained 81% accuracy.

Similarly, Jayaprakash et al. (2020) proposed the random forest algorithm and an in-house variation, the improved random forest algorithm. These algorithms yielded 91 and 93% accuracy, respectively, when predicting academic performance.

In their work, Daud et al. (2017) used the C4.5 and CART algorithms to forecast whether students would complete their coursework. Through these algorithms, the authors achieved 76.6 and 71%, respectively, in F1-score.

Ma et al. (2018) proposed using deep neural networks to predict the passing rate of online students. They used data from Guo Pei Online Education. First, the features that most affect passing rate were filtered through the filter-type feature selection algorithm, which provided 27 representative features. Then, preprocessing and standardization were performed on the remaining dataset (27 features). The decision tree algorithm was used for prediction, and the grid search algorithm was used to optimize the decision tree. The following metrics were used with these algorithms: precision, recall, F1-score, and runtime. Before testing, the data were divided into pass and fail. Finally, using the grid search algorithm at 50% pass / 50% fail data, decision tree achieved 96% accuracy.

Of all the literature in this section, six authors used decision tree algorithms: Kumar et al. (2020), Castrillón et al. (2020), Daud et al. (2017), Ma et al. (2018), Hasan et al. (2019), and Amazona and Hernandez (2019), while Jayaprakash et al. (2020), Rincon-Flores et al. (2020), and Benablo et al. (2018) used random forest algorithms. Despite random forests being a subset of decision trees, most studies use decision trees for multiple-variable predictions, such as those for academic performance.

In their study, Hasan et al. (2019) present a model that tries to predict final exam results for a given student. For this, they used a dataset of 1,170 students in 3 courses. Then, the authors preprocessed the dataset by removing unnecessary columns such as Student ID. They used the KNN algorithm and a decision tree classifier (the ID3 algorithm) for their predictions, thereby obtaining 94.44% accuracy based on the decision tree classifier algorithm.

In their research, Benablo et al. (2018) forecast student performance based on social network data, such as Facebook, Twitter, Instagram, and YouTube, as well as online games. A single-format input was created to serve as the training dataset following the attribute relationship file format. To finalize the model, the researchers considered using support vector machine (SVM), KNN, and random forest for prediction. Furthermore, to ensure that the model accurately classified a given dataset, a 10-fold cross validation was performed using the WEKA software. For these modes, three metrics were used: precision, recall, and F1-score. For the results, the authors used three types of samples (30, 50, and 100 instances). After testing 100 instances, random forest achieved a result of 100% in precision, 80.6% in recall, and 89.3% in F1-score.

Amazona and Hernandez (2019) proposed to use a decision tree for prediction following the same methodology as Kumar et al. (2020) and Ma et al. (2018). The results from the decision tree were 93% in precision, 96% in F1-score, and 100% in recall.

From Kumar et al. (2020), Castrillón et al. (2020), Daud et al. (2017), Ma et al. (2018), Hasan et al. (2019), and Amazona and Hernandez (2019), Rincon-Flores et al. (2020) reported the highest accuracy rate using a decision tree algorithm at 96%. Moreover, it considered the second-highest number of forecasting variables (27).

The study by Blanco et al. (2016) focuses on predicting student results for the Data Structures I and II classes because these computer engineering classes present a significant degree of difficulty and require serious dedication and rigor from students, which means that their grades are often lower than desirable. Therefore, the authors designed a model based on the MATLAB system that predicts academic results in these classes after having been trained with specific data from each class. The model’s architecture is designed with three layers, wherein an activation function was used for each neuron: for the input and hidden layers, the sigmoidal hyperbolic tangent function was used. For the output layer, the linear function was used because the authors wanted to achieve the largest possible range amplitude in the output interval, which also facilitated result interpretation. In this study, a prediction effectiveness of over 78% was achieved for the first subject and 75% for the second subject.

The research reported by Sekeroglu et al. (2019) was developed based on the need to improve AI-based systems in the field of education because in-class and remote students often need help to improve their performance. Therefore, two neural network models were used: backpropagation (BP) and long/short-term memory (LSTM). The former uses a gradient descent algorithm during learning and propagates the error to update weights and minimize error values; the latter memorizes previous neural network inputs to provide more accurate results. The results were as follows: BP (70% variance) and LSTM (77.9% variance).

Amazona and Hernandez (2019) also proposed using a deep learning neural network model for prediction. The same methodology as above was used. The deep learning results were as follows: 98% precision, 97% F1-score, and 98% recall.

Of the studies carried out by Blanco et al. (2016), Amazona and Hernandez (2019), and Sekeroglu et al. (2019), the neural network that reported the highest accuracy was the one reported by Amazona and Hernandez (2019) at 98% precision, 98% recall and 97% F1-score. Moreover, Amazona and Hernandez (2019) used the least number of input variables at only 9.

In this sense, Blanco et al. (2016) used the highest number of input variables, 21. Nevertheless, its effectiveness rate when predicting student performance was considerably low (75% accuracy) compared with the effectiveness reported by Amazona and Hernandez (2019) (95% accuracy). It may be observed that Ma et al. (2018) uses fewer layers and, therefore, fewer network neurons, which leads to a more inaccurate process.

Daud et al. (2017), used a supervised learning model to predict whether students will complete or abandon their study programs. Specifically, they used the SVM model, wherein the best result was obtained with 86% in the F1-score test.

Ma et al. (2018) also used the SVM supervised learning model to predict online student passing rates. Using the grid search algorithm at 50% pass / 50% fail data, this model achieved 95% accuracy.

Sekeroglu et al. (2019), used the SVR model to improve AI-based systems in the field of education. This model provided predictions at 79.7% variance.

Benablo et al. (2018) forecasted student performance based on social network data, such as Facebook, Twitter, Instagram, and YouTube, using the SVM model. In a 100-instance test, this model reported 100% precision, 96.8% recall and 98.4% F1-score, thereby being the most effective model.

In their study, Liao et al. (2019) sought to predict which students are at risk of achieving poor performance in a given class. This must be identified early enough to allow instructors to help students before they fall behind. Thus, they designed a model that predicts the student’s final exam grade using a binary SVM classifier, which is trained with the radial basis function kernel based on the selected course. Additionally, the model is based on three parameters that indicate the severity of different classification error types. Based on this model, at least 62% of the at-risk students were correctly identified.

From Daud et al. (2017), Ma et al. (2018), Sekeroglu et al. (2019), Benablo et al. (2018), and Liao et al. (2019), the supervised learning model that reported the highest accuracy was the one used by Benablo et al. (2018) at 100% precision, 96.8% recall, and 98.4% F1-score. To reach this accuracy, the authors used cross validation 10 times to avoid overfitting the model. Here, we can see a pattern: Sekeroglu et al. (2019) did not use cross validation, and it reported the lowest accuracy rate at 79.7%. Similarly, Daud et al. (2017) performed cross validation but only five times, which is half as many times as Benablo et al. (2018). Daud et al. (2017) reported an accuracy rate of 86%, which is considerably high but not as high as the accuracy rates reported by Liao et al. (2019), Ma et al. (2018), and Benablo et al. (2018), who performed the cross validation 10 times.

This section will summarize and compare the most relevant models in the literature, which will be shown in Table 1. Data such as the technique used, accuracy, precision, recall, etc. were analyzed. It should be noted that not all models had all validation metrics.

According to Alaieri and Vellino (2017), where different ethical regimes are applicable, artificial intelligence decisions must be reliable and duly justified so that users, manufacturers, and legislators can understand how decisions are made and what ethical principles were applied in each case. This study presents a model that breaks down the stages of ethical decision-making into their elementary components. The proposed project references ethical decision-making matters regarding the algorithm used to predict student performance. The ethical issues are detailed in Section 3.2.

Figure 1 illustrates the architecture used in the model proposed.

To track the evaluation of the training and validation process of the artificial neural network, the Accuracy and Loss metrics were used in each epoch.

Figure 8 illustrates the Accuracy and Loss recorded for each epoch. On one hand, Training Accuracy, in orange, reaches a certain percentage where it remains stable, and Validation Accuracy, in red, follows the same behavior as Validation Accuracy. On the other hand, Loss Training and Validation follow a similar decreasing pattern until reaching a minimum percentage, where they stabilize.

A total of 100 epochs were required for training in the hyperparameters. However, to avoid overfitting, EarlyStopping was used, thereby cutting off at 54 epochs upon detecting that the accuracy percentage remained constant.

The model achieved 93.81% accuracy, which indicates that the prediction was correct for the percentage of cases validated. Likewise, 94.15% precision was obtained. This measures the quality of the model, indicates the number of students who actually passed the course, and predicts that they would pass the course. In other words, it is the percentage of students correctly identified by the model as passing from the total number of students identified by the model as passing. The model achieved 95.13% recall, which indicates the number of approved students that the model identified, that is, the percentage of students correctly identified as approved from the total number of approved students. Finally, the model obtained an 94.64% F1-score, which is the combination of the accuracy and recall metrics as a single value.

This section compares the performance of the proposed model with the models that had the best performance according to the literature. The most relevant studies are specified in Table 1. Three machine learning models were implemented, which were trained with the same set of datasets with which the neural network was trained in order to compare them.

According to Benablo et al. (2018), who evaluated the SVM and KNN techniques, the SVM model provides 98% accuracy due to the type of model used to predict performance as a function of the input data. Benablo et al. (2018) also details three other models used in which a variation in the instances of the dataset is seen. The higher the number of instances, the higher the accuracy, sometimes reaching 100%. After the implementation of our own SVM model, and the comparison with the proposed neural network model, we can observe which neural network is better both in accuracy and in the test results shown in the Table 3.

After comparing with the Amazona and Hernandez (2019) proposal, three different models were used: Naïve Bayes, deep learning and decision tree, with an accuracy of 96, 98 and 93%, respectively. We opted to implement our own Naïve Bayes model to train it on our dataset for comparative purposes. The results of the Naïve Bayes model were the lowest in the comparisons performed, as can be seen in Table 3.

In his research, Rodrígiez-Hernández et al. (2021) implemented an artificial neural network to predict the academic performance of senior students. The research notes that they used information from a sample of 162,030 students and that the trained model provided an accuracy of 82%. However, they also used personal information about the students (e.g., their socioeconomic status, household characteristics, personal background, the types of schools they had attended, and the salary of working students) to train the model. In this case, to perform the comparison of the models with the same data set, the Rodrígiez-Hernández et al. (2021) model would have to be modified, so a correct comparison cannot be performed; however, we can observe that 8 of the 10 attributes used as input data, are personal data, this means that if we remove the personal data from their model, the model would suffer a decay in performance. Yağcı (2022) implemented several models to predict students’ academic performance.

The models used were random forest, neural network, SVM, logistic regression, Naïve Bayes and KNN, where the neural network provided the best results with an accuracy of 86.30%. Yağcı (2022) did not use students’ personal data. They only used their midterm and final exam grades, career path and academic department as information. The 86% accuracy obtained by this research study is the highest accuracy rate. Thus, a proprietary Random Forest model was implemented, trained with the same data set with which our proposed model was trained. The result of the Random Forest model was the best performing among the implemented models, this can be seen in Table 3.

The research of Yousafzai et al. (2021) compares the existing models, such as RNN, CNN, LSTM, SVM and BiLSTM, with the model proposed by the authors, which achieved a better accuracy of 90.16%. Yousafzai et al. (2021) uses personal data, such as age, gender, address, parents’ occupation, type of school and parents’ educational level, to train the model. The dataset used for training is mainly composed of personal data. Similar case to Rodrígiez-Hernández et al. (2021), due to the architecture developed by the authors and the database used, it is not possible to perform an equivalent comparison; however, we can observe that from the model proposed by Yousafzai et al. (2021), 15 of the 33 entries for his model are personal information of the students, such as: age, location, love relationship, father’s jobs, mother’s jobs, etc. which means that if we were to remove those 15 entries, the model would suffer a loss of performance.

Comparing our model with the literature, we can observe that personal data significantly affect the model results, regardless of the type of model used. For this reason, and due to the increasing vulnerability of personal data evidenced by data breaches and leaks, we propose a student performance prediction tool that does not expose sensitive student and teacher information and yields better results than the observed literature.