Studies on the predictions of student performance generally use demographic, academic, and behavior characteristics to predict Grade Point Average (GPA), course grade writing Graduate Management Admission Test (GMAT) or Scholastic Assessment Test (SAT) scores (if available), or graduation status (Nedeva & Pehlivanova, 2021). Several ML techniques are applied: Logistic Regression (LR), Naïve Bayes (NB), Decision Trees (DTs), Neural Networks (NNs), and Support Vector Machines (SVMs), as well as boosting algorithms (Al-Shanableh et al., 2024). Existing research has indicated that the performance of different learning system and learning models are subject to source data size, feature type and model selection.

Shahiri, Husain, & Rashid (2015) conducted a review of EDM studies during 2002–2015 and observed that DTs and Neural Networks had the highest accuracy rates (of up to 98%). Another study by Namoun and Alshanqiti (2020) analyzed 586 research papers and found that the RF and Hybrid Neural Networks performed better than classical statistical techniques. Albreiki, Zaki, & Alashwal (2021) conducted a survey of 78 studies and discovered that: classification algorithms, such as SVMs, RF and NB, are often used in practice (especially for at-risk students’ prediction) (Al-Shanableh et al., 2024).

ML is further divided into three types: single, ensemble and hybrid (Várkonyi-Kóczy, 2020). Single models are based on a single classifier, and ensemble models integrate such homogeneous models through bagging, boosting or stacking to strengthen generalization. Hybrid models combine several mixed techniques and typically include optimization or feature selection (Al-Shanableh et al., 2026).

Recent works have shown that hybrids of classifiers do better than single loners, especially if we are dealing with complex data. Kumar, Singh, & Handa, (2017) reported an increase of 75.62% in accuracy by combining Radial Basis Function (RBF) and Multi-layer Perception (MLP) NNs with respect to single models. Accordingly, using 4-algorithm hybrid classifier, Sokkhey and Okazaki (2020) obtained an accuracy between 84.9 and 99.7%. However, few works have been published that study hybrid methodologies for higher-education databases in MiddleEastern contexts.

Al-Husban (2021) used several algorithms in Jordan and employed a data set from Al al-Bayt University to predict student status and obtained 77% accuracy with XGBoost as the best single model among others. Extending this work, Mashagba (2022) revealed that CatBoost was the best model (92.16%) for predicting student academic status with boostingbased algorithms. Nonetheless, none of these works used a hybrid ensemble approach.

Hybrid and ensemble learning models are of great interest to researchers. Both ensemble models and the hybrid model use the integration principle but in a simple twist, while ensemble ML integrate homogenous models, the hybrid classifier integrates heterogeneous models (Al-Shanableh et al., 2026).

A grouping model combines models to make a group decision for prediction. The hybrid classifier takes more features for filtering; the reason for calling it hybrid is that it learns about data pre-processed flow and model building, in contrast to ensemble there are no constraints on data processing. In Hybrid ML, one classifier classifies each model (Wong & Yeh, 2020).

Although existing literature has proved the superior performance of Boosted Decision Tree Classifiers for education prediction tasks, little effort has been devoted to hybrid ensemble classifiers that combine CatBoost with RF and XGBoost together. Besides, only a few works use large real-world datasets from Middle East contexts and only a limited number of studies compare hybrid vs single models on the same dataset. The current paper fills these gaps through a comparative study that includes up to 19,700 student results from Al alBayt University and a newly developed hybrid softvoting classifier.

Albreiki, Zaki, & Alashwal (2021), in their research “A Systematic Literature Review of Student’ Performance Prediction Using Machine Learning Technique”, performed a systematic review of EDM literature over the period 2009 to 2021 (78 studies reviewed). The most common type of datasets used in studies consisted of the ones from student university databases and online learning platforms. Data mining significantly increased student achievement because it was very successful in predicting at-risk students and dropout rates.

Albreiki, Zaki, & Alashwal, (2021)’s literature review showed 16 research studies into predicting student performance, 12 research studies into recognizing those at risk, and five studies to determine how e-learning affects student academic achievement. The most popular methods used were Decision Tree (DT), Logistic Regression (LR), Naive Bayes (NB) and Support Vector Machine (SVM). Student’s drop-out prediction (21 papers) was the second most common metatask, with the primary integrated methods being Decision Tree (DT), Support Vector Machine (SVM), Classification and Regression Trees (CART), K-Nearest Neighbors (KNN), and Naive Bayes (NB). Twenty-four studies focused on predicting students’ performance using static and dynamic data (14 studies using a combination of methods) where the algorithms most commonly used were KNN, NB, SVM, DT, RF I.D3 & ICRM.

Namoun and Alshanqiti (2020) demonstrate this in “Predicting Student Performance Through Data Mining and Learning Analytics Methods – A Systematic Literature Review.” This is the one adopted in Bunkar et al. (2020), when they used some criteria to select papers, and left out what did not accept them. They began with 586 articles and filtered it down to 62 papers “to adjudicate.” The most utilized learning types were statistical analysis with 28 articles, supervised ML with 25 articles, and data mining with five articles. The distribution of algorithms/learning models were:

We dichotomized the models into best versus worst. The best performing models were 3L Neural Network (feedforward) 98.8%, RF 98%, Hybrid RF 99%, Naive Bayes 96.8% and ANN (95–97%). The least accurate models were linear/cox regression (50%) and logistic regression (76.2%), repeated discriminant analysis (64–73%), mixed effect logistic models’ analysis (69%), and bagging (48–55%).

Bunkar et al. (2020) used Clustering, Classification and Association Rules in E-learner. In clustering, the most widely used algorithm that addresses the solution to a problem is K-Means clustering algorithm. The classifier was Apriori algorithm. The algorithms employed for association rules have been J48, C4.5, REPTree, and Naïve Bayes.

Aldowah, Al-Samarraie and Fauzy (2019) conducted a literature review from 2000 to 2017. 402 articles were included in the review. They found 26% used classification methods, followed by 21% using clustering, 15% using visual data mining, 14% using statistics, 14% using association rule-mining, and 10% using regression. Romero and Ventura (2007), on the other hand, reported a different outcome in their review, finding that association rule mining was employed over that for classification (43% vs 28%) and clustering (15%). This is in line with the work of Papamitsiou and Economides (2014), who also applied a large number of classification methods effectively prior to clustering and regression.

The work of Ashraf, Anwer & Khan (2018) also compared data mining methods and classification algorithms in terms of the impact they have on datasets attributes’ influencing student performance predictions (see Figure 3).

In another vein, Shahiri, Husain, & Rashid (2015) indicated that in their investigation “The Third Information Systems International Conference 2065 A Review on Predicting Student’s Performance using Data Mining Techniques.” During 2002 – 2015, they found ten papers that used DT as learner to predict student performance, whilst eight other papers utilized NN algorithms; four used Naïve Bayes algorithm; and lastly three papers employed K-Nearest Neighbor (KNN) and SVM for predicting the output variable. They proved that for these algorithms, the best prediction accuracy of all was achieved by NN (98%), and the second best is DT (91%). For the KNN and the SVM, accuracy was equated (83%). Naive Bayes is in the third place with 76% accuracy.

The authors used SVM algorithm most frequently, following the work of Rastrollo-Guerrero, Gómez-Pulido, & Durán-Domínguez (2020), and found it to be the best at making predictions. In addition to SVM, DT, NB and RF are common algorithmic recommendations that have been well-investigated and yielded positive results. Even if neural networks are not widely used, they seem to be excellent at predicting academic achievement of pupils (Rastrollo-Guerrero, Gómez-Pulido, & Durán-Domínguez, 2020).

The work of Al-Husban (2021) proceeded in the same line. Using application/implementation analysis with real data, the study applied to collected concerning outcomes of Al-al-Bayt University Jordan. This amounts to a dataset of 25,017 students. Al-Husban developed a model to predict different majors (Graduate/Non-Graduate) of the students by categorizing the students in binary classes. She uses multiple classifiers as well, such as XGBoost, RF, SVM, KNN, and DT for her algorithms. The data was split into a 75% training set and a 25% test set for all functions. She concluded by predicting with the following metrics: 2.4 Using Accuracy, Precision, Recall and F1 Score for power generation. The result shows XGBoost classifier has the better accuracy (of 77%) than all others.

Moreover, Al-Husban has completed a model for predicting the students’ performances in two scenarios – predicting whether a student will succeed or fail, andpredicting students’ level of appreciation. The thesis of Mashagba (2022) was used as it contained actual data collected from Al-al-Bayt university in Jordan (it is very close to our study’s proposed use cases of data, particularly that having same aim). More than three algorithms of Gradient Boosting were applied to this dataset including AdaBoost, CatBoost, and XGBoost. A 10-fold cross-validation was worked against grid-search, looking for optimal split point and combine parameter value. The experimental results indicated that the CatBoost algorithm has better prediction accuracy, with 92.16% for the final status predicting model and 86.89% for the appreciation prediction model, respectively. She calculated the meta-tracker performance for the following metrics: Accuracy-, Precision-, Recall-, and F1-scores.

In terms of predicting student performance, we observed that most algorithm types (NN, DT, NB, SVM – and even RF and LR) were used. Information about students’ demographics and grades, including their grades in different studies and their high school qualifications, behavioral data, Moodle access modules, personal details (i.e., gender), and academic performance are being employed to predict successfulness of students. Table 1 shows a summary of previous research and the most used approach for students’ performance prediction.

The first dataset is a sample of the AABU student population that contains academic and demographic attributes for 19,700 students from many faculties and departments across all academic years. For prediction model development, the clean file remained after data preparation (removal of cases with missing and duplicate records). The synthetic dataset was constructed using ChatGPT (GPT-4) with the following pipeline: A structured human-friendly prompt was conceived and input to Chat GPT and the task at hand was promoted: show us what a realistic student record would look like, which includes the same features as the AABU datase. Demographic distributions were given (age 18–25, gender ratio of 55:45 female to male), academic performance ranges (GPA between 50 and 100 with normal distribution μ=70, σ=12), and high school rates (between 60 and 99). The synthetic data were tested against the real AABU dataset in terms of distributional statistics (mean, standard deviation, correlation matrices) to ensure realistic representations. We used this dual-dataset approach: a) to assess the model generalizability across different data sources, b) to deal with privacy issues by releasing a shareable complimentary fake synthetic dataset and c) to investigate the robustness of our models against differences in data generation processes.

Features for both datasets are described in Table 2. The dependent variable is the student academic performance, which is divided into excellent performance, very good, just good, passing and failing. The aim is to accurately group students according to the appropriate outcome category from these features.

Data preprocessing included: (1) Handling of missing data: Records with more than 30% of missing values were removed, and remaining missing values were imputed using mode for categorical attributes and median for numerical ones; (2) Encoding of categorical attributes was applied by ordinal encoder (for ordered categories like rating levels) or one-hot encoding (for nominal categories like specialization and gender); (3) Dropping irrelevant and privacy-sensitive columns such as student name, national ID, phone number, email; (4) Outlier detection by Interquartile Range method, where we flagged any value outside 1.5 times IQR; (5) Normalization via Min-Max scaling transformation of numerical columns to ensure same scale across algorithms; and finally (6) Balancing classes using SMOTE (Synthetic Minority Over-sampling Technique).

Some additional features were constructed on the base of the original ones. These were annual means, year of admission categories and cumulative scores. Selection of features was done using recursive feature elimination (RFE) and information gain (IG). This approach improved model as well processing time.

The following ten single ML classification algorithms were evaluated: LR, NB, DT, KNN, SVM, RF, Gradient Boosting, XGBoost, Adaptive Boosting (AdaBoost), and Categorical Boosting (CatBoost) All experiments were implemented using Python 3.9 with scikit-learn 1.2.0, XGBoost 1.7.0, AdaBoost 3.3.5, and CatBoost 1.1.1. Key hyperparameters were optimized using GridSearchCV with 5-fold cross-validation: Logistic Regression (C=1.0, solver=’lbfgs’, max_iter=1000); DT (max_depth=10, min_samples_split=5); Random Forest (n_estimators=100, max_depth=15); KNN (n_neighbors=5, metric=’euclidean’); SVM (kernel=’rbf’, C=1.0, gamma=’scale’); XGBoost (n_estimators=100, max_depth=6, learning_rate=0.1); CatBoost (iterations=500, depth=6, learning_rate=0.1); AdaBoost (n_estimators=50, learning_rate=1.0)). Data was split into 80% training and 20% testing sets with stratified sampling to maintain class distribution.

The hybrid model consists of three models:

The choice of these models was based on individual model performance and the type of the classifier. Both Soft Voting Ensemble (Averages probability predictions) and Hard Voting Ensemble (Majority vote) was tried.

Performance was evaluated based on Accuracy, Precision, Recall and F1-score. The performance metrics used in this study are defined mathematically as follows: Accuracy = (TP + TN) / (TP + TN + FP + FN), measuring the proportion of correctly classified instances; Precision = TP / (TP + FP), measuring the proportion of true positive predictions among all positive predictions; Recall (Sensitivity) = TP / (TP + FN), measuring the proportion of actual positive instances correctly identified; F1-Score = 2 × (Precision × Recall) / (Precision + Recall), the harmonic mean of precision and recall providing a balanced performance measure. Where TP = True Positive, TN = True Negative, FP = False Positive, and FN = False Negative. The equations are also shown in Figure 5. Moreover, 10-fold cross-validation was used to improve the reliability of evaluation.

The data were collected in an official way through university administrative staff with the consent of The Dean of Student Affairs. To ensure ethical treatment of student related information, we took the following measures: (1) We removed all identifiable student data such as student names, national identification numbers, phone numbers, email addresses and home addresses from the dataset before any analysis can be done; (2) anonymous unique identifiers were generated through an irreversible cryptographic hash function to replace real Student IDs for de-identification in order to protect re-identification of individual students; (3) The files were kept on secure encrypted university servers utilizing password protection to restrict access exclusively to members of the research team; (4) Prior to even looking at the data, all research team members signed agreements ensuring confidentiality and non-disclosure; Additionally, Data handling procedures are designed based on FERPA act and Jordanian laws regarding use and transfer of educational record and personal identifiable information that protects researchers’ access or misuse of educational Personal Information with JPR. For the synthetic case, no ethical approval was necessary because the data are synthetic and do not contain subjects.

In the first step, an individual ML algorithm was evaluated within a 10-fold cross validation. The accuracy of all the single models is shown in Table 4. Overall, best performing with the highest accuracy on the AABU and synthetic datasets was XGBoost (Accuracy 80% and 79% respectively) as shown in Tables 4 and 5. Figure 7 shows the accuracy comparison.

The summary of experiment results shows that among the single classifiers, XGBoost obtained the highest accuracy rate (80%), followed, very closely, by Random Forest. Classic algorithms such as SVM, Naïve Bayes and KNN demonstrate less accuracy in terms of precision because it is difficult for them to deal with multi-dimensional data and complex and nonlinear relationships.

Hybrid Ensemble Model was created by combining Random Forest, Logistic Regression and XGBoost, using soft voting and hard voting. Combination of the various classification pairs (RF+ LR+XGBoost), demonstrated highest predictive degree, 92.06 points one-day ahead with all-AUC above 90% in any case. For the hybrid ensemble model in our research, the result was better than all single classifiers. Comparative results their performance is shown in Table 6.

When multiple strong learners are combined, model performance and misleading rate indicate that the benefits of hybrid models lie in their robustness (Al-Shanableh et al., 2026). The soft voting solution provides a stepping-stone for hybrid models to use probability-based decision fusion to enhance generality.

Overall, the findings of this study suggest that our combined model could predict future performance of students much more accurately than any one single classifier had done in previous research. These results not only demonstrate the potential of real-life institutional data; they are also a kind of quality assurance and guardrail to lean on for automatic academic risk detection systems. Our research shows that a mixed ML model has better accuracy and stability than the classification model in predicting the academic performance of a student.

The best overall accuracy (92.06%) was achieved with our hybrid model, which combined XGBoost, Logistic Regression and Random Forest. This was compared to CatBoost, the best single classifier (80%). Furthermore, the hybrid had improved robustness across validation folds and less misclassification, particularly in low academic achievement categories. These conclusions are consistent with the findings of previous studies on the effectiveness of ensemble and hybrid learning methods in complex, heterogeneous educational data (Kumar, Singh, & Handa, 2017; Sokkhey & Okazaki, 2020). The hybrid model integrated multiple strong learners into a single system, leveraging beneficial decision boundaries while dampening these turning points in such a way that generalization was also promoted and algorithmic bias eased. Soft voting achieved further enhancements by probability-based aggregation: the combination of strong learners could be managed even better than voted Majority Rule or Not.

Feature importance analysis identified seven significant influential factors for academic performance. The cumulative GPA, high school average, academic year in high school and course load were the only important predictors of student college GPA. These predictions are also consistent with those in similar research, which also discusses the impact of students’ academic history and background variables on their performance. Schools rely on accurate prediction of students’ academic behavior to help them develop reliable plans. However, by identifying at-risk students in time, it can offer more targeted academic advising, and lead to better resource allocation; such measures can encompass the use of personalized learning or interventions that reduce dropouts and increase the effectiveness of the whole institution. Predictive tools that take data as an input and turn it into information in university management systems help universities to plan for their future, plan intelligently, and plan well.

This research proves the superiority of ensemble learning in educational data analytics. Hybrid systems like these combine the abilities of several ML models (for example, robustness and strong modelling on research in multi-dimensional space) with all those advantages that single classifier cannot bring off, such as dealing with non-linear data and heavily imbalanced categories. Today we might say that “the most essential combination” based on the above analysis builds up around “enslaving processes wirelessly”, which are envisaged as yielding superior results. The result concerning the importance of features was that, across algorithms, academic history variables (for instance, cumulative GPA high school code program concentration which academic year) are key predictors, confirming similar findings from education data mining studies.

Although the findings are encouraging, there are several constraints that should be considered. First, the present study is conducted based on a single university (AABU); thus, findings may not be easily generalized to other institutional contexts of student populations, academic systems or grading styles. Second, even though the synthetic dataset is suitable for validation purposes, it might not be able to replicate all of the complexity and subtleties present in actual student data, which can have an impact on the reliability of cross-dataset comparisons. Third, the study is limited to academic and demographic factors; behavioral measures, including class attendance, learning management system (LMS) use, library utilization, and extracurricular involvement were unavailable but could increase predictive ability. Fourth, the historical patterns used in the analysis are assumed to hold in the future, which won’t be true during times of great institutional changes or exogenous shocks (e.g., large-scale transitions to online learning related to a pandemic). Fifth, although the performance of the hybrid model was superior to all models considered herein, its computational complexity may be too high for real-time use in resource-limited institutional settings. Future studies should address these limitations by: (1) externalizing the model between different institutions with different cultural and educational settings; (2) extending realistically oriented behavioral and engagement features; (3) investigating model interpretability to generate meaningful feedback for educators, and (4) creating lightweight model variants ready to be deployed on education information systems.

In sum, these results confirm the value for educational communities and AI researchers alike of incorporating hybrid classification models in student performance prediction.