Using machine learning for early identification of At-risk students

Abstract

The use of machine learning (ML) in education offers promising opportunities to enhance teaching strategies and improve student outcomes. Predictive models can identify at-risk students early, enabling timely and targeted interventions. This study applies ML techniques to predict students’ academic performance using a real-world dataset from a higher education institution. The dataset consists of 1419 student records with 12 features describing the academic attributes. The preprocessing steps includes handling missing values, outlier removal, data standardization, and feature engineering to improve data quality and model performance. Six regression models including linear regression, ridge regression, decision tree regressor, random forest regressor, gradient boosting regressor, and support vector regressor were evaluated after comprehensive preprocessing and model optimization. Among these, the Random Forest Regressor achieved the highest predictive accuracy with R² equals 0.93. These results highlight the potential of regression-based approaches for deployment in early-warning systems, supporting data-driven decision-making and improved educational outcomes.

Introduction

In recent years, data mining and related technologies have transformed numerous fields, including marketing, healthcare, and education, as they enable the discovery of meaningful patterns within vast datasets^1,2. The rapid development of data mining methods, particularly through artificial intelligence (AI) and machine learning (ML), has driven significant innovations across various domains, reshaping how information is analyzed and applied³. In education, these advancements have spurred interest from researchers and institutions, as data mining techniques increasingly inform the ways we understand and enhance teaching and learning practices⁴. Many higher education institutions (HEIs) have started implementing these tools to optimize both academic and administrative processes, recognizing their potential to improve educational outcomes and institutional effectiveness^5–7.

From an academic standpoint, data mining enables educators and administrators to tailor instruction and resources to the specific needs of students, enhancing their educational experience. Techniques such as classification and clustering are employed to optimize resource allocation, customize learning materials to individual learning styles, and adapt the difficulty of exams and assignments based on student profiles^8,9. Moreover, data mining supports categorizing learning materials based on expected outcomes, improving how content is aligned with student goals^10,11. Prediction models have become a focal application of these techniques in education, with HEIs increasingly using them to forecast academic performance, student retention, and enrollment trends^12–15.

Despite these advancements, challenges remain in fully integrating data mining and predictive models within educational settings. Developing accurate predictive models that account for diverse learning patterns, social backgrounds, and academic behaviors is complex. While prediction models can provide early warnings for at-risk students, identifying which variables significantly impact academic success is still difficult due to the variability in student performance and external influences such as report presented in¹⁶. Nonetheless, data mining in HEIs offers a robust approach for educators and policymakers to understand student strengths, weaknesses, and learning patterns, thereby supporting data-driven decisions that directly impact educational practices.

Educational Data Mining (EDM), a term coined in the early 2000s¹⁷, focuses specifically on using data mining techniques to address educational challenges. EDM extracts valuable insights from datasets generated within educational settings, including academic records, attendance, assessments, and even socio-demographic factors. These insights allow institutions to identify hidden patterns that influence student success and to make proactive interventions. Through AI and machine learning, EDM has become instrumental in analyzing data from various sources, including learning management systems (LMS) such as Moodle, Blackboard, Canvas, and MOOCs, to improve student support and predict academic outcomes^18–22. The COVID-19 pandemic accelerated LMS adoption and increased the volume of digital records, enhancing the availability and quality of educational data. These datasets can include information on demographics, academic performance, class participation, Grade Point Average (GPA), attendance, and learning styles, providing a foundation for models that enhance teaching and learning environments as reported in¹⁸.

The first year of university is particularly crucial in shaping students’ academic trajectories. Success during this period lays a foundational understanding of essential knowledge and skills, influencing performance in subsequent years. First-year achievements can also boost students’ confidence, helping prevent dropouts and changes in major. In addition, GPA requirements for scholarships and admission to competitive programs mean that early academic success can significantly impact students’ future opportunities. Thus, predictive models can help identify students who might struggle early in their academic journey, enabling institutions to offer targeted support to improve retention and overall academic performance.

This study aims to explore machine learning techniques for predicting student performance, with the goal of contributing to the development of early-warning systems for students at risk of failure. By investigating predictive models in the context of educational data mining, this research seeks to provide valuable insights that could shape educational policies, improve teaching strategies, and ultimately support students in achieving academic success.

In this study, an at-risk student is defined as a student whose predicted academic performance indicates a strong likelihood of failing or achieving a grade below the institutional passing threshold. Although the proposed models estimate academic outcomes as continuous values through regression techniques, these predictions can be translated into risk categories according to institutional grading policies. For instance, students whose predicted final grade falls below the passing threshold may be categorized as high risk, whereas those with predicted grades near the threshold may be considered at moderate risk. This approach allows regression models to function as early warning tools while maintaining flexibility for institutions to establish their own intervention criteria.

Overall, the main contributions of this study are:

• Design and implementation of machine learning-based predictive models to identify students at risk of academic failure early in the semester, allowing timely interventions.

• Comparison of multiple machine learning techniques (e.g., Linear Regression, Ridge Regression, Decision Tree Regressor, Random Forest Regressor, Gradient Boosting Regressor, Support Vector Regressor) to identify the most accurate and efficient models for predicting student performance.

• Generation of actionable insights into factors that influence academic success, sup- porting data-driven decisions that can inform educational policies and curriculum design.

• Enabling educators to identify students who may benefit from additional resources or support, thereby enhancing personalized learning approaches and boosting retention rates.

• Establishing a framework for applying machine learning models to educational datasets, which can be adapted and expanded upon in future studies to improve predictive accuracy and applicability across diverse educational contexts.

To guide the investigation, this study focuses on three research questions. First, how accurately can regression-based machine learning models predict student academic performance using commonly available academic indicators? Second, which regression model provides the most reliable predictive performance for the early identification of students at academic risk? Third, how can regression-based predictions be applied to support early-warning systems and institutional intervention strategies in higher education?

The structure of the paper is outlined as follows: Sect “Literature review” explores existing literature related to educational data mining. Section “Academic performance prediction model” elaborates on the proposed methodology for the operation of utilizing classifications models and regression models. Section “Results” provides an evaluation of the proposed solution through experimental results. Section “Discussion” offers a discussion on utilizing classification and regression models within the context of education. It also highlights some limitations experienced during this study. Finally, Sect. “Conclusion and future work” concludes the study along with the future directions.

Literature review

Starting with Artificial Neural Network (ANN), researchers revealed that ANN has the ability to detect dependency and interaction between dependent and independent features extracted from dataset²³. Researchers considered data related to psychometer factors, cumulative grade point average (CGPA), demographic data, and school background. For instance, the study presented in²⁴ used ANN algorithm and confirmed that there is a strong influence of fundamental courses to predict engineering students’ CGPA. A prediction model proposed in²⁵ has considered Multilayer Perceptron Neural Network (MPNN) algorithm to predict students’ performance based on features: interest of study, unit test grade, assignments’ grade, attendance score, extracurricular activities, and residence place. Datasets were collected for real data for 300 students. A 10-fold cross validation is used for the training data and the accuracy of obtained results was 91%. Researchers in⁶ proposed a model utilized ANN which considered students’ navigation statistics through LMS.

In term of Support Vector Machine (SVM) algorithm, researchers in²⁶ investigated the use of perceptron classifier, support vector machine and decision trees, logistic regression and random forest for predicting students’ academic progress to classify students to three classes (failing, low passing grade, and high marks). The research work showed that SVM achieved 70.8% accuracy results on predicting students’ performance based on dataset contains 480 records. Each record has features related to demographic data, grades, study year level, and students’ engagements such as number of raised hands, how often students check announcements, number of accessed resources, etc.

A number of prediction models were proposed to utilize decision tree due to its simplicity and easy conversion to IF THEN rules²⁷. Different studies were performed to identify the attributes and factors that can help decision tree algorithm in predicting both teachers and students’ performance^28,29. For instance, some of the proposed attributes and factors can be extracted from log data files in educational web-based websites and platforms^18,30. Particularly, the attributes are related to final grades, accumulative GPA, final grades in some particular courses. Researchers in³¹ showed the use of decision tree in identifying students with high risks of dropout of college. The study depended on features that are related to gender, attendance, previous semester grade, educational level for parent, parent income, if the student is first child, having a scholarship, and working or not. Other researchers investigated the use of decision tree combined with other techniques such as ANN³².

Other algorithms were utilized to forecast academic performance. For instance, Gradient Boosting Machine (GBM) were considered in³³ to predict students’ academic outcomes at the end of the school year. The proposed model considered 17 features from the dataset including 238,575 records and 247,297 records in the years 2015, and 2016 respectively. The results showed the validation data Receiver Operating Characteristic (ROC) curve was more than 89% depending on the dataset. Hybrid approach was also thoroughly investigated in the literature. For instance, authors in⁷ considered dataset including the midterm exam grades, final exam grades, faculty, and department of 1854 students. The utilized algorithms were the random forests, nearest neighbor, support vector machines, logistic regression, Naïve Bayes, and k-nearest neighbor algorithms. The study showed a comparison between the previous algorithms to predict students’ final grades. The study achieved a classification accuracy more than 70% after applying all previous mentioned algorithms.

More overview on types of features considered in the prediction models and data mining techniques adopted to predict students’ performance in the body of literature can be found in^34–37. For instance, the study³⁸ compared five different machine learning algorithms to identify features and attributes that can affect the prediction of students’ academic performance. The algorithms were multinomial logistic regression, artificial neural network, random forest, gradient boosting and stacking methods. The study considered features related to three categories: socio-demographic category (obesity, average household income, family size, etc.), academic category, and student-related category (stress and lifestyle). The results showed gradient boosting algorithm obtained best results among the other algorithms. Similar to the previous study³⁹, presented a comparative study for seven different prediction methods to identify at-risk students. Logistic Regression, Support Vector Machine, Decision Tree, Multi-Layer Perceptron, Naive Bayes Classifier, K-Nearest Neighbor and Ensemble model (SVM, K-Nearest Neighbors, and Naive Bayes Classifier). The adopted dataset in this study included grades for exams and quizzes, weekly homework, and team participation during one semester. The results showed that Naive Bayes Classifier model and an Ensemble model outperformed results among the other investigated methods.

In light of previous studies, developments and potential challenges, this paper contributes to the body of knowledge by presenting a comprehensive analysis of six specific regression techniques. These techniques carefully chosen to represent different groups of regression techniques based on their characteristics for predicting students’ performance, the chosen techniques include traditional linear models such as Ordinary least Squares (OLS) Linear Regression⁴⁰ and Ridge Regression⁴¹, which offer wide interpretability and simplicity in regression, as well as ensemble method techniques such as Decision Tree Regressor which support continuous and categorical features⁴², Random Forest Regressor⁴³, and the Gradient Boosting Regressor⁴⁴ which useful in capturing complex relationship and handle non-linearities in the dataset⁴². Additionally Support Vector Regressor is considered in the proposed diverse regression approach for its effectiveness in handling high dimensional data.

Many prior studies on student performance prediction employ classification models that assign students to predefined categories such as pass or fail, or low and high risk^5,39,45. While these approaches offer clear and actionable labels for intervention, they may overlook variations in the degree of academic difficulty among students. Regression models, by contrast, generate continuous predictions of academic performance. This allows institutions to assess not only whether a student may be at risk but also the extent of that risk. As a result, regression-based predictions can support more refined early-warning systems, where the type and intensity of intervention are adjusted according to the predicted level of performance rather than relying solely on binary classification.

Most research on predictive learning analytics has been conducted in universities in North America and Europe, where large educational datasets and well-developed digital infrastructures are readily available. In contrast, institutions in the Global South often operate under different conditions, including limited data availability, smaller datasets, and varying levels of institutional resources¹³. Recent studies in regions such as the Middle East, North Africa, and Latin America have begun to explore the use of machine learning models for predicting student performance within these contexts¹³. These efforts underscore the need for predictive approaches that rely on commonly available academic indicators and that can be implemented effectively in resource-constrained educational environments. Table 1 shows a summary of related work based on datasets, adopted ML methods in the study, prediction type whether it is classification or regression and obtained results.

Table 1.

Summary of Related Studies on Student Performance Prediction.

Study	Dataset / Context	ML Method	Prediction Type	Key Findings
Delen (2010)12	University dataset	Decision Trees, Neural Networks	Classification	Demonstrated effectiveness of ML models for predicting student retention and dropout.
Nandeshwar et al. (2011)14	University student records	Pattern learning models	Classification	Identified behavioral patterns linked to student retention.
Arsad et al. (2013)22	Engineering students dataset	Artificial Neural Network	Regression	Achieved strong performance in predicting student CGPA using academic indicators.
Romero et al. (2008)25	LMS activity data	Decision Tree	Classification	Demonstrated ability of data mining to classify student learning behavior.
Marbouti et al. (2016)36	Course-level dataset	Naïve Bayes, SVM, Decision Tree	Classification	Developed early-warning models to detect at-risk students during the semester.
Fernandes et al. (2019)31	Public school dataset (Brazil)	Gradient Boosting	Regression	Achieved strong predictive accuracy for academic performance using socio-demographic and academic variables.
Waheed et al. (2020)6	LMS big data	Deep Learning	Regression	Showed that deep learning models can capture complex learning patterns.
Yağcı (2022)7	University dataset	Random Forest, Logistic Regression	Classification	Random Forest achieved the best performance for predicting student success.
Rajendran et al. (2022)35	High school dataset	Gradient Boosting, Random Forest	Classification	Ensemble models outperformed traditional classifiers.
Marcolino et al. (2025)28	Moodle log dataset	ML optimization models	Classification	Showed predictive analytics can detect dropout risks using LMS interaction data.
Malik et al. (2025)34	Educational datasets	Ensemble ML models	Classification	Feature selection improves prediction accuracy significantly.
Proposed Study	University dataset (1419 records)	Random Forest, Gradient Boosting, SVR, Linear Models	Regression	Random Forest achieved R² = 0.93, demonstrating strong predictive capability for student academic performance.

Results of this study will support educational policies and practices, enabling the educational institutions to make data-driven decisions that enhance the learning experience and ultimately contribute to the academic success of students. By leveraging these findings, educational stakeholders can proactively implement effective interventions tailored to students’ needs as well as enhancing the learning experience.

Academic performance prediction model

In this research, similar to previous studies^28,29 the adopted methodology includes integrated techniques and machine learning tasks such as data preprocessing, feature engineering, and model validation and evaluation. This adopted approach is designed to achieve more accurate predictions for students’ performance at any stage of their academic journey. Therefore, this research sought to identify patterns and factors that influence student performance in higher education institutions. The model is constructed based on a comprehensive real dataset obtained from Arab American University in Palestine. The experimental protocol was approved by the university too. The obtained students’ records were anonymized to follow up relevant guidelines and regulations. The proposed prediction model consists of three main phases: Data Preprocessing, Implementation of Prediction Model, Results Evaluation and Discussion. The subsequent paragraphs present more details for each phase to provide a comprehensive understanding of the adopted research methodology in this research work.

Data preprocessing

The preprocessing of data is essential for ensuring model accuracy and generalizability. The following steps outline the mathematical approach for each preprocessing task:

Handling missing values

Let D be the dataset with n samples and m features. For each feature X_j, missing values are handled by either removing rows with excessive missing data or imputing values. For imputation, let x_ij represent the i-th value in feature X_j. If x_ij is missing, it is replaced by the mode mode(X_j):

1

Outlier removal

Outliers are identified based on a threshold method or statistical approach such as using Z-scores. For each data point xij, calculate the Z-score:

2

where µ_j and σ_j are the mean and standard deviation of feature X_j. If |z_ij| > threshold, x_ij is considered an outlier and removed or replaced.

Data standardization

Standardization is performed on each numeric feature X_j to scale values within a range of [0, 1] using Min-Max scaling. Let x_ij be the i-th data point of feature X_j. The standardized value x^′_ij is calculated as:

3

where min(X_j) and max(X_j) represent the minimum and maximum values of X_j.

Feature engineering

The dataset contains categorical features that are encoded using label encoding. For a categorical feature Xk with unique categories C1, C2,. . ., Cr, assign each category unique.

numeric label:

Label(Ci) = i, i = 1,2,…,r (4).

This transforms categorical features into a format usable by machine learning algorithms. Table 2 shows the features that have been extracted from the dataset.

Table 2.

Adopted Features from the Dataset.

Features	Mean	STD
Middle-term-score	0.5156	0.1890
Participation score	0.7176	0.1782
Attendance rate	0.9346	0.1167
School average	0.5717	0.2134
Final exam	0.5181	0.2045
GPA score	0.5857	0.1793
Final grade	0.4618	0.2164
Gender	0.4017	0.4904
Major program name	0.6283	0.2755
School certificate type	0.5187	0.4895
School branch	0.8809	0.1630
Final Result	0.4973	0.2609

To enhance reproducibility, the dataset was split into training, validation, and testing subsets using a 70–15–15% stratified split, ensuring that the distribution of key performance variables was preserved across all partitions. The training set was used for model fitting, the validation set for hyperparameter tuning through grid search and cross-validation, and the testing set for final performance evaluation. Prior to model training, all features underwent exploratory statistical analysis to assess their distribution, central tendency, and variability. For example, attendance rates exhibited a high mean (0.93) with low variance, while midterm and final exam scores displayed moderate variability, indicating their potential discriminative power in predictions; as shown in Table 2. Categorical variables such as gender and school branch were analyzed for balance to avoid skewed representation. Outlier detection was conducted using Z-scores, and extreme values beyond the threshold of ± 3 standard deviations were examined and either removed or imputed. This systematic preprocessing approach was performed to ensure that the dataset was clean, representative, and suitable for robust model training and evaluation.

Prediction model implementation

This section describes the formal structure of the proposed prediction system, including the system inputs, preprocessing modules, learning models, and the mathematical formulations used for model training and evaluation.

The dataset D was loaded and organized using the Pandas DataFrame structure in Python⁴⁶. For this research, all computations were performed on an HP EliteBook 840 G2 with 8 GB of RAM and an Intel Core i7 CPU at 2.60 GHz, running Microsoft Windows 10 Pro Education (64-bit). To ensure robust model evaluation, the dataset was split into training and testing subsets. Specifically, D was partitioned into a training set Dtrain (70%of the data) and a testing set Dtest (30% of the data), such that:

D = D_train ∪ D_test, where |D_train| = 0.7 × |D| and |D_test| = 0.3 × |D|.

The prediction model was constructed using a diverse regression approach, wherein each regression model M ∈ { M1, M2,., Mk } is associated with a set of hyperparameters θ to optimize its performance. To find the optimal hyperparameters θ∗ for each model, an exhaustive grid search technique, implemented via the GridSearchCV method⁴⁷, was employed. This approach evaluates each model M over a specified hyperparameter space Θ to minimize a chosen error metric (e.g., Mean Squared Error, MSE).

The hyperparameter optimization process is defined as follows:

where n is the number of samples, represents the loss function, yi is the true label, and M(xi; θ) is the prediction made by model M with hyperparameters θ.

Furthermore, cross-validation was employed to ensure unbiased evaluation and selection of the best model configuration. This process involves partitioning Dtrain into k-folds and averaging the model performance across each fold. The optimal model with the best-performing hyperparameters was selected based on cross-validated error metrics, yielding a robust model configuration ready for evaluation on Dtest.

The model evaluation phase is essential for assessing how accurately the model performs on newly generated data. To determine the regression model that provides the most accurate predictions, three key evaluation metrics are calculated and reported: Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared (R2).

• Mean Squared Error (MSE): This metric measures the average squared difference between the predicted values and the actual values, indicating how close the predictions are to the actual outcomes. MSE is calculated as follows:

where n is the number of data points, yi represents the actual value of the target variable for the i-th data point, and yi′ represents the predicted value for the i-th data point.

• Mean Absolute Error (MAE): MAE computes the average of the absolute differences between the predicted values and the actual values. This metric provides insight into the average error magnitude without considering direction. It is calculated as follows:

• R-squared (R²): R² is a commonly used metric to evaluate the predictive accuracy of regression models. It represents the proportion of variance in the actual values that is explained by the predictions. R² is calculated as follows:

where y ¯ is the mean of the actual values yi. An R2 value close to 1 indicates a strong explanatory power, while a value near 0 suggests weaker predictive performance.

Each of these metrics provides distinct insights into model performance, allowing for a comprehensive evaluation of accuracy and reliability.

Algorithm 1 describes a machine learning approach for predicting student performance by utilizing a structured dataset containing academic and demographic features. Data preprocessing is first conducted to handle missing values, standardize numerical data, and encode categorical attributes. Key features are then selected through recursive feature elimination, and additional informative variables, such as the final grade, are computed. The dataset is split into training and testing subsets, after which six regression models (Linear, Ridge, Decision Tree, Random Forest, Gradient Boosting, and Support Vector Regressor) are trained, with hyperparameters optimized using GridSearchCV. Model accuracy is evaluated based on Mean Squared Error, Mean Absolute Error, and R-squared metrics, and actual versus predicted values are visualized to assess each model’s performance. The predictions are intended to inform educational interventions for students at risk of under-performing, with suggestions for future work involving expanded datasets and additional models for even greater predictive accuracy.

Algorithm 1.

Student performance prediction model.

The selection of the six regression models including Linear Regression, Ridge Regression, Decision Tree Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Support Vector Regressor was guided by the need to represent a diverse spectrum of machine learning approaches with varying strengths. Linear Regression and Ridge Regression were included as interpretable baseline models that provide a benchmark for predictive accuracy and enable straightforward evaluation of feature effects. Decision Tree Regressor was selected for its ability to handle both numerical and categorical data while producing easily interpretable decision rules. Random Forest and Gradient Boosting Regressors were chosen as widely recognized ensemble techniques known for their robustness, capacity to model non-linear relationships, and high predictive accuracy in educational data mining tasks. The Support Vector Regressor was included for its effectiveness in high-dimensional spaces and its ability to generalize well with limited data points. The models enable a comprehensive comparative analysis across simple, regularized, tree-based, ensemble, and kernel-based methods, ensuring that the findings are not limited to a single algorithmic family and providing a balanced foundation for both performance evaluation and practical deployment.

Regression models were selected because they generate continuous estimates of student performance rather than categorical labels. This enables educators to assess the degree of academic risk and tailor the intensity of interventions accordingly. In addition, continuous predictions can be readily converted into categorical risk levels by applying institutional thresholds, offering flexibility for implementation under different educational policies.

As many early-warning systems adopt categorical risk outputs (e.g., low, medium, high) to guide interventions, this study deliberately employed regression models to predict continuous performance scores. Continuous predictions offer greater granularity, allowing educators to interpret not only whether a student is at risk but also the degree of risk, which can inform the intensity and type of intervention. Furthermore, continuous outputs can be post-processed into categorical risk levels using institution-specific thresholds, providing flexibility for deployment across diverse educational contexts without retraining the predictive model. This dual utility supports both precise academic forecasting and practical risk classification, aligning the model’s outputs with a range of institutional needs.

Results

Systematic experiments were conducted on the dataset introduced in Sect. “Data preprocessing” to assess and compare the effectiveness of the selected machine learning models in predicting student performance. The configuration and performance of each model, along with their fine-tuned hyperparameters, are detailed in Table 3. These parameters were carefully selected using grid search and 5-fold cross-validation to optimize performance.

Table 3.

Hyperparameters tuning of the proposed ML models.

Model	Hyperparameters
Linear Regression	-
Ridge Regression	Alpha value: 10.0
Decision Tree Regressor	max_depth: 5
Random Forest Regressor	n_estimators: 200
Gradient Boosting Regressor	n_estimators: 50
Support Vector Regressor	Kernel: rbf, C: 10.0

The evaluation results highlight the models’ predictive accuracy through Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared (R2) metrics as presented in Table 4. The MSE values ranged from 0.00 to 0.01 across all models, indicating minimal deviation between predicted and actual outcomes. This low MSE confirms that models such as Linear Regression, Ridge Regression, Decision Tree Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Support Vector Regressor effectively minimize prediction errors. The MAE results further underscore model accuracy, with the Random Forest Regressor achieving the lowest MAE of 0.05.

Table 4.

The performance of the proposed ML models.

Model	MSE	MAE	R-squared
Linear Regression	0.01	0.08	0.88
Ridge Regression	0.01	0.08	0.88
Decision Tree Regressor	0.01	0.07	0.86
Random Forest Regressor	0.00	0.05	0.93
Gradient Boosting Regressor	0.01	0.07	0.91
Support Vector Regressor	0.01	0.08	0.87

R-squared values reveal high explanatory power, particularly for the Random Forest Regressor, which attained an impressive R2 of 0.93, capturing 93% of the target feature variability. The Gradient Boosting Regressor also performed robustly, with an R2 value of 0.91, closely following the Random Forest model. These results demonstrate the models’ precision and efficacy in capturing essential patterns within the data.

To verify whether the differences between the compared ML models are statistically significant, statistical hypothesis testing was conducted using 10 cross-validation results. First, a Friedman test was applied to compare the performance of all evaluated models using the R² metric. The results indicate statistically significant differences among the models (p < 0.01) as depicted in Table 5.

Table 5.

Wilcoxon Signed-Rank Test for Random Forest compared with Other Regression Models.

Comparison	p-value	Significant (α = 0.05)
RF vs. Gradient Boosting	0.002	Yes
RF vs. Linear Regression	0.002	Yes
RF vs. Ridge Regression	0.002	Yes
RF vs. SVR	0.002	Yes
RF vs. Decision Tree	0.002	Yes

In order to analyze these differences, the Wilcoxon signed-rank test was conducted using the Random Forest Regressor as the control model. The results show that the Random Forest model significantly outperforms Linear Regression, Ridge Regression, Decision Tree, Gradient Boosting, and Support Vector Regression (p < 0.05). These results confirm that the superior performance of the Random Forest model is statistically significant rather than the other models due to random variation across folds.

Figure 1 illustrates the performance of the models using scatter plots to represent the relationship between actual and predicted values. In these plots, the blue dots indicate actual student performance values, while the red dashed line serves as a reference for an ideal prediction scenario where actual and predicted values align perfectly. The scatter plots reveal that while the model predictions deviate somewhat from this ideal line, they capture the general trend of student performance, with a noticeable concentration of points near the diagonal line, suggesting an overall strong correlation. The evaluation results clearly demonstrate that the ensemble models, particularly the Random Forest and Gradient Boosting Regressors, exhibit highly accurate predictive capabilities for student academic performance compared to individual linear and regularized regression models. These results suggest that ensemble techniques significantly enhance both the accuracy and explanatory power of regression models in this educational context, providing a valuable approach for more precise performance predictions.

Fig. 1.

Visualization of actual vs. predicted values using six regression models.

Discussion

The performance of the proposed framework, particularly the Random Forest and Gradient Boosting Regressors, aligns with and in some cases exceeds the accuracy levels reported in prior studies on student performance prediction. As highlighted in the literature, ensemble-based methods have consistently demonstrated strong predictive capabilities in educational data mining, often outperforming simpler linear models and single decision trees when applied to academic datasets as shown in recent studies on ensemble learning and feature selection for educational prediction tasks^37,48. While differences in dataset characteristics, institutional contexts, and feature sets make direct one-to-one comparisons challenging, the high R² value of 0.93 achieved in this study suggests that the proposed approach performs competitively with state-of-the-art methods documented in recent research. This reinforces the suitability of ensemble regression models as a robust foundation for early warning systems, which offers both predictive accuracy and flexibility for adaptation to diverse educational settings. This answers first and second research questions.

While this study focused on the application and comparison of established machine learning algorithms, this choice was to create a validated baseline framework for predicting student performance in higher education. The dataset, drawn from a single mandatory course and consisting of commonly available academic variables such as grades, attendance, and participation, was selected to ensure a consistent assessment context across diverse student year levels and to facilitate practical deployment of the model in institutional settings. Although this scope limits the generalizability of the, it aligns with the objective of providing an immediately implementable early-warning system that requires minimal additional data collection^39,45. Future work will address these limitations by incorporating multi-course, multi-semester datasets enriched with demographic, behavioral, and engagement features, and by exploring methodological innovations such as interpretable AI models, optimized ensemble learning, and deep learning architectures tailored to educational contexts. This incremental approach ensures both the practical relevance of the present work and a clear pathway toward broader applicability and methodological advancement.

As this study aims to identify the most effective regression model for predicting students’ academic performance based on a dataset from a mandatory course, there are several limitations to consider. First, demographic information, such as parental status, living area, and marital status, was not included due to data unavailability, potentially limiting the model’s predictive scope. Additionally, the analysis focused exclusively on a mandatory course, which may not fully capture variations in student performance across elective courses with different difficulty levels and engagement requirements. This focus on an obligatory course was intended to ensure a sufficient sample size across various student year levels (Freshman, Sophomore, Junior, Senior), yet extending this research to elective courses could yield further insights.

It is also important to note that while many early-warning systems adopt categorical risk outputs (e.g., low, medium, high) to guide interventions, this study deliberately employed regression models to predict continuous performance scores. Continuous predictions offer greater granularity, allowing educators to interpret not only whether a student is at risk but also the degree of risk, which can inform the intensity and type of intervention⁵. Furthermore, continuous outputs can be post-processed into categorical risk levels using institution-specific thresholds, providing flexibility for deployment across diverse educational contexts without retraining the predictive model⁴⁵. This dual utility supports both precise academic forecasting and practical risk classification, aligning the model’s outputs with a range of institutional needs.

It is also worth mentioning that the deployment of predictive models in educational settings must account for issues of fairness, interpretability, and ethical use of student data⁴⁹. Algorithmic bias, if unaddressed, can disproportionately affect students from underrepresented or disadvantaged backgrounds, potentially reinforcing existing inequalities⁵⁰. Techniques such as balanced training datasets, bias detection metrics, and stakeholder review of model outputs are critical to mitigating such risks⁵¹. Equally important is the interpretability of models: in educational contexts, decision-makers including teachers, advisors, and administrators must understand the rationale behind predictions to ensure transparency, trust, and meaningful interventions. This consideration often favors models or post-hoc explainability techniques (e.g., SHAP values, LIME) that translate complex algorithms into actionable insights^4,5,12. The ethical and legal aspects of handling sensitive student data demand strict adherence to privacy frameworks, informed consent procedures, and institutional data governance policies. In the regional context, this requires compliance not only with global best practices but also with local regulations, ensuring that predictive analytics serves the educational mission without compromising students’ rights.

From a computational perspective, the models were trained on a relatively small dataset (1,419 samples with 12 features), and the experiments were conducted using standard computing resources with low training time. This indicates that the framework can be implemented in typical institutional environments without requiring specialized hardware. However, when applied to larger educational datasets, ensemble models such as Random Forest and Gradient Boosting may require additional computational resources, although they can scale effectively using parallel or cloud-based computing environments¹³.

Beyond predictive accuracy, the proposed framework has practical relevance for institutional policy and academic support strategies. Integrating the predictive model into institutional systems, such as student information systems or learning management platforms, would allow universities to monitor key academic indicators throughout the semester^20,28. When predicted performance falls below predefined thresholds, automated alerts could notify academic advisors or instructors, enabling timely interventions such as tutoring, counseling, or personalized learning support. In this way, the framework serves not only as a predictive model but also as a decision-support tool that helps safeguard students’ educational trajectories and reduce the risk of academic failure or dropout⁵². This answers the third research question raised in the beginning of the study.

Conclusion and future work

In this experimental study, a unique dataset on student academic performance from obligatory course for students in Arab American University was processed and analyzed. Through extensive preprocessing steps, including normalization, handling of missing values and outliers, and feature selection, six regression machine learning models were tested: Linear Regression, Ridge Regression, Decision Tree Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Support Vector Regressor. The results indicated that ensemble models, particularly the Random Forest and Gradient Boosting Regressors, demonstrated superior predictive capabilities compared to individual linear and regularized models. These findings suggest that ensemble techniques significantly enhance the accuracy and explanatory power of models in predicting student performance.

The insights from this research have practical implications for educational institutions and stakeholders. By leveraging ensemble models to predict academic performance in mandatory courses, educators can gain a deeper understanding of factors impacting student outcomes, enabling them to design more targeted interventions and personalized learning strategies. Integrating this predictive model into a university’s portal or LMS could provide real-time progress updates to identify at-risk students, facilitating timely support and optimized teaching strategies to promote academic success.

The practical implementation of the proposed predictive framework depends on several institutional conditions. Universities must first ensure reliable access to timely academic indicators, including attendance records, assessment scores, and participation data, which constitute the primary inputs for the predictive models. In addition, model outputs should be integrated into institutional platforms such as learning management systems or student information systems to enable continuous monitoring of student progress throughout the semester. Institutions must also establish appropriate performance thresholds to determine when students are considered at risk and to activate corresponding intervention protocols. Moreover, effective use of such systems requires human oversight, whereby academic advisors and instructors interpret the predictions and provide appropriate support. When these conditions are met, predictive analytics can operate as a practical early-warning mechanism that informs institutional decision-making and supports improved student outcomes.

Based on this research, several future directions are proposed. From a practical application perspective, future work should examine the scalability of the proposed models to larger, multi-institution datasets and assess their robustness across different academic contexts, course types, and student populations. Attention should also be given to data governance, including privacy protection, ethical handling of sensitive student information, and compliance with relevant regulations, to ensure responsible deployment in real-world educational environments. Longitudinal evaluation of model performance over multiple semesters would help determine stability, adaptability to curriculum changes, and sustained predictive value. In addition, investigating which specific features most influence predictions could enhance interpretability and trust in the models. Methodologically, expanding the analysis to include robustness analysis of the models and other alternative machine learning techniques, such as Neural Networks and Naïve Bayes, as well as deep learning approaches, may further improve accuracy and provide richer insights to support targeted, timely interventions that enhance student success.

Author contributions

Conceptualization, A.E. and S.D.; methodology, A.E., S.D. and A.A.; validation, A.A, S.D.; formal analysis, A.A, HA.A. and HI.A; investigation, Y.H., HA.A. and A.A.; sources, A.E.; data curation, A.A.; writing—original draft preparation, A.E and A.A.; writing—view and editing, Y.H., A.E and HI.A; visualization, A.E, S.D.; supervision, A.E., and Y.H.; funding acquisition, Y.H. All authors have read and agreed to the published version of the manuscript.

Funding

The authors did not receive any direct financial funding for this research.

Data availability

The datasets generated during the experimentation of this study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Informed consent

The requirement for informed consent was waived, as the dataset contained no identifiable information.