Regularized multi-path XSENet ensembler for enhanced student performance prediction in higher education

Introduction

Due to limited resources, educational institutions worldwide continue to face the substantial problem of student retention (Mduwile & Goswami, 2024). Institutional intervention tactics have shifted their attention to early detection of at-risk students, especially during their first academic year, due to the high average dropout rate of 49% in OECD nations (Wijnia et al., 2024). Machine learning (ML) and data mining techniques are widely used across various domains, including healthcare, marketing, and finance, to support predictive analysis (Mohzana, Murcahyanto & Haritani, 2024; Liu, Heath & Grzywacz, 2024; Li, Wang & Wang, 2024; Abulibdeh, Zaidan & Abulibdeh, 2024; Dawar et al., 2024; Gong et al., 2024). Educational data mining (EDM) applies these methods in the education sector to academic data, helping detect patterns, personalize learning, identify disengagement, and support timely interventions (Baek & Doleck, 2023).

Despite the rise of predictive tools, the growing complexity of academic data presents challenges in performance assessment, especially when considering sustainability goals (Martins et al., 2024). ML has proven effective in forecasting student outcomes and identifying at-risk learners, yet more emphasis is needed on predictions using pre-admission data. Recent advancements have expanded this landscape: personalized SPOCs have improved English as a foreign language (EFL) learning (Jiang et al., 2024), gamified systems have enhanced behavior (Li & Jianxing, 2025), and meta-universe platforms and innovative education policies have supported adaptive learning (Chen, Jin & Chen, 2024; Liu, Cao & Chen, 2024). Studies on MOOC engagement (Liu et al., 2024) and socio-environmental factors, such as parenting and resilience (Cao et al., 2024), further underscore the need for integrating technology, policy, and psychology into predictive models.

Choosing the proper prediction technique remains complex due to the diverse range of hyperparameters and data contexts. Improved ML (IML) addresses this by automating model and parameter selection (Li et al., 2024). However, its empirical application in educational settings remains limited. Our work bridges this gap by applying IML for early-stage student performance prediction, utilizing data available at the time of academic entry. This study contributes to academic management by analyzing performance trends through the proposed XSEJNet model. The framework ensures fairness, transparency, and integration readiness, providing consistent and accurate forecasts. The core contributions are summarized as follows:

1. This study presents XSEJNet, a scalable and interpretable machine learning framework designed to overcome the limitations of manual and traditional school management systems. The model is tailored for educational datasets, with careful consideration of computational complexity.

2. To ensure robust prediction of student performance, XSEJNet addresses key challenges such as class imbalance, missing data, and biased distributions. The study introduces standardized modeling practices that improve prediction consistency across diverse educational contexts.

3. A core emphasis of the proposed model is on accessibility and explainability, enabling educators and stakeholders to clearly interpret the factors contributing to student success or underperformance.

4. XSEJNet integrates advanced feature representation techniques to enhance both accuracy and computational efficiency, making it suitable for real-world educational deployment and decision-making.

5. The findings provide practical insights that can inform more effective teaching strategies, student support initiatives, and institutional planning—benefiting educators, universities, and policy-makers alike.

Related work

Over many years, academic research has focused on issues related to student success. Different studies have investigated the factors that affect educational outcomes. The study used neural networks, decision trees, and support vector machines to analyze students’ internet use habits and academic achievement (Viveka & Priya, 2024). The use of neural networks and student registration data for pattern detection led to the development of a prediction framework (Masadeh & El-Haggar, 2024). In order to identify students who may be at academic risk and to provide timely interventions and support, recent studies have integrated bidirectional long short-term memory (BiLSTM) networks with logistic regression (Hopcan, Türkmen & Polat, 2024). Aiming to improve academic support systems and decrease performance discrepancies, recent projects have used deep learning to identify learning patterns and provide personalized help (Judijanto, Atsani & Chadijah, 2024; Stojanov & Daniel, 2024). Natural language processing (NLP) and other language-based artificial intelligence approaches have the potential to improve student learning, according to recent developments in educational technology. In order to find language patterns that might be a predictor of academic performance, researchers examine student writing in Almazroi, Shen & Mohammed (2019). Administration operations, such course scheduling, have been improved with the use of data mining (Martins et al., 2024; Monsalve-Pulido et al., 2024). Furthermore, systems that suggest courses based on each student’s unique academic history and goals have been created with the aim of boosting engagement and retention rates (Almazroi et al., 2016). Smarter and more responsive educational environments are created by data-driven technology and machine learning. Shirawia et al. (2024) stresses the need of complex algorithmic models that are both highly performing and easily scalable. Furthermore, the authors proposed using hierarchical evaluation methodologies and rubrics to gauge the effectiveness of education (Pallathadka et al., 2023). Academic programs are customized to satisfy real-world demands via industry-education collaborations. A number of studies have looked at new ways to identify children who could be at danger. According to Zhao et al. (2022), one approach to identifying challenging individuals in both undergraduate and graduate programs made use of tree-based models and genetic fuzzy systems. In parallel, teacher performance was assessed using several categorization methods, with the C5.0 algorithm demonstrating superior accuracy and reliability (Luo, Han & Zhang, 2024). An analysis of decision tree models revealed that the Random Tree method outperformed J48 in predicting third-semester academic performance (Zhai et al., 2024). New developments utilize hybrid and deep learning to enhance accuracy and increase explainability. For example, the reinforcement learning co-evolutionary hybrid intelligence (RLCHI) model in Vimarsha et al. (2024) combines reinforcement learning with expert educator input, resulting in better performance than RL alone. The Enhanced AEO-XGBoost model, optimized with metaheuristics, exhibited good accuracy with Support vector machine-Synthetic Minority Oversampling Technique balancing (Cheng, Liu & Jia, 2024). In Alshamaila et al. (2024), a convolutional deep learning model addressed class imbalance in big datasets using oversampling and undersampling. According to Huang & Zeng (2024), dual graph neural network (DualGNN) is a graph neural network that excels at interaction-based and attribute-based learning, resulting in good categorization.

In Rainio, Teuho & Klén (2024), researchers evaluated skill competency based on academic success and knowledge scope to improve prediction and classification algorithms. The research compared decision tree, ensemble, and label algorithm methods to determine the most effective classification strategy. This study enables educational institutions to identify and address students’ ability gaps, thereby enhancing their learning experiences and long-term outcomes. Researchers highlighted the need to develop educational data mining for sustainable insights and improved student performance. Using tree-based algorithms, ML predicts academic results and identifies at-risk pupils, linking assessments to performance. However, unknown places exist. Quality assurance algorithms for sustainable higher education and inclusive features for various students demand attention. These results support the application of ML in academic administration, promoting egalitarian and environmentally friendly education. The existing literature is presented in a structured manner in Table 1.

This research demonstrates that combining educational experience with artificial intelligence creates more accurate, transparent, and adaptable prediction models.

Proposed methodology

The objective of the proposed framework is to create a stable method for forecasting academic achievement by seamlessly integrating the standard for data mining, which encompasses comprehensive student performance data. This ensures that the model operates efficiently. Additionally, various data mining techniques focus on categorization to minimize resource consumption and promote responsible data analysis practices. Figure 1 shows the system model’s architectural depiction.

The analysis of student performance and best practices used the Student Educational Performance (SEP) dataset (Aljarah, 2024; Amrieh, Hamtini & Aljarah, 2016; Liu et al., 2022; Cohausz et al., 2023). This study uses this diverse educational dataset. Many components indicate learning processes. The information includes student demographics, academic status, interactions with learning resources, and engagement. This wealth of data lets academics study educational relationships, patterns, and trends. EDM applications benefit from the comprehensiveness of the SEP dataset. It helps researchers model learning, enhance education, and understand student behavior. The data is preprocessed to remove missing values and outliers, ensuring data integrity. The biases and errors are decreased, which may affect the classification model. A two-step feature selection method was employed to identify categorization features. The first approach utilized the extreme gradient boosting (XGB) algorithm, which effectively identifies crucial features and intricate linkages. This method revealed critical parameters affecting student achievement prediction accuracy. In step two, the feature set is increased using recursive feature elimination (RFE). RFE enhances model generality and interpretability by removing minor traits. Furthermore, substantial exploratory data analysis (EDA) is conducted to gain a deeper understanding of the dataset’s complexity. Variable distributions and relationships, including categorical features, are examined. This extensive study examined signals and trends that improve the classification system.

Classification criteria are identified using XGB and RFE algorithms, focusing on variables that impact student success. EDA is needed to improve categorization. Category data are numbered to ease processing. Data is split into 30% test evaluation and 70% training data to maintain fairness. Parameter space is explored using the Jaya algorithm (JA) to improve model performance. This optimization strategy dramatically enhanced forecast accuracy. The novel classification system is compared to several existing approaches in the literature to test its durability. This rigorous research analyzed the long-term advantages and excellence of the academic classification approach. Several literature models are used, including XGB, support vector machine (SVM), naive Bayes (NB), K-nearest neighbor (KNN), residual neural network (ResNet), bidirectional encoder representations from transformers (BERT), extreme gradient boosting-tree-based optimization tool (XGB-TPOT), and SVM ensemble models. This method was tested for accuracy, area under the curve (AUC), precision, logarithmic loss, F1-score, and memory use. This work rigorously evaluated several prediction models. Statistical investigations use variance analysis, significance testing, and other measures to ensure accuracy. Considering resource restrictions and computational complexity, these stringent tests demonstrated the proposed method’s effectiveness. The XSEJNet ensemble model outperformed competing models on all assessment criteria, demonstrating its accuracy in forecasting student success.

Proposed XSE method

The XSE model represents an intricate and adaptable architecture for forecasting student performance. The above framework combines the methodical learning and contextual grasp provided by the squeeze-and-excitation (SE) attention model with the powerful feature extraction capabilities offered by the residual networks with external transformations (ResNeXt) architecture. Figure 2 shows how the XSE Model is structured internally. The process starts with student performance data being passed through several parallel paths based on the ResNeXt architecture. Each path carries out a series of operations—convolution, batch normalization, and rectified linear unit (ReLU) activation—to extract meaningful features from the input. These features are then combined through concatenation, allowing the model to retain diverse representations due to the multiple paths. After this, the combined features enter a SE based parallel modeling layer. Here, the model uses gated mechanisms to understand patterns over time by learning from past performance states. This helps the model focus on the most relevant information while filtering out less valuable signals, ultimately improving its ability to predict student outcomes over time.

ResNeXt mechanism of extracting features (Li, Zhu & Zhu, 2023): Initial pupil achievement data is included in the prediction procedure. Within the ResNeXt component, this data undergoes several modifications. Steps for each route i in the ResNeXt block are as follows during the feature extracting stage:

• Layer_Convolutional: $Con{v}_i={\mathrm{L}\mathrm{a}\mathrm{y}\mathrm{e}\mathrm{r}\mathrm{\_}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{o}\mathrm{l}\mathrm{u}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_i(x)$

• Normalization of Batches: $B{N}_i={\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{z}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{\_}\mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}}_i(Con{v}_i)$

• Trigger Activation using ReLU Function: $ReL{U}_i=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}(B{N}_i)$

In this case, the input student performance data is denoted by x. The variables $Con{v}_i$, $B{N}_i$, and $ReL{U}_i$ denote the outputs of the convolutional layer, the outcome following the batch normalization procedure, and the resulting result learning the activated ReLU function to the group-normalized output for path i.

Concatenation and the principle of cardinality: The data was systematically divided into several routes according to cardinality to provide a model with the ability to gather various attributes. The ResNeXt architecture processes the information separately along each path. Several pathways’ outputs, which are enhanced with unique data, are carefully combined to produce a complete feature depiction:

(6) $$\mathrm{P}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{\_}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=[fun{c}_{R}eLU1,fun{c}_{R}eLU2,fun{c}_{R}eLU3,fun{c}_{R}eLU4]$$The concatenation operator combined all pathways’ outputs ( $fun{c}_{R}eL{U}_i$) (four in this example).

SE-based parallel modeling: After the concatenation phase, the combined features pass via the squeeze-and-excitation-based parallel modeling layer. This layer allows the model to analyze sequential relationships and detect temporal dependencies in the student accomplishment data while maintaining the ResNeXt architecture.

(7) $$a=\alpha \left(M_a\cdot \left[\mathrm{P}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{\_}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n},b_{t-1}\right]+c_a\right)$$

(8) $$b=\alpha \left(M_b\cdot \left[\mathrm{P}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{\_}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n},b_{t-1}\right]+c_b\right)$$

(9) $${\tilde{c}}_t=\beta \left(D\cdot \left[\mathrm{P}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{\_}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n},a\cdot b_{t-1}\right]+e\right)$$

(10) $$c_t=(1-b)\cdot b_{t-1}+b\cdot {\tilde{c}}_t$$

These formulas elucidate the operational mechanics of the SE-based parallel modeling layer within the ResNeXt framework, where ‘c_t’ signifies the concealed state during time ‘t,’ and gates ‘a’ and ‘b’ regulate the data flow through adaptive scaling processes.

Simulation results

This research utilized Python on a Core i7 processor with 16 GB of RAM. This analysis selects an extensive dataset from Aljarah (2024), including many factors related to student achievement. Table 3 demonstrates that the proposed dataset has diverse features that provide superior insights over previous research.

This dataset was extensively evaluated for its suitability in relation to the research aims and academic usage before being selected. Python was chosen for its data analysis skills, vast library, and flexibility. Data processing was efficient, utilizing a Core i7 CPU and 16 GB of RAM, which made analytical procedures and modeling straightforward.

Figure 3 (Nationality) additionally divides into the number of persons by country of origin. According to the data above, China has the highest count, and Jordan has a higher frequency than other nations. To reach these conclusions, the study’s dataset was examined for its nationality variable. Calculating value counts and percentages enables the determination of the distribution of various nationalities. This data was visually shown using a bar plot for easy visualization. Figure 3 (Topics) depicts the frequency distribution of several subjects and the related statistics from educational institutions in various countries. The picture shows that the quantity of IT, French, and scientific themes is substantially higher than that of other courses.

Table 4 presents an ablation study of the XSEJNet model, analyzing how each component contributes to its overall performance. The complete model achieves the best accuracy, F1 score, and log loss results. When SE attention or ResNeXt is removed, performance drops notably, showing their role in refining and extracting deep features. Excluding Jaya optimization increases training time, and skipping feature selection leads to poor generalization. These findings confirm that each module plays a vital role in XSEJNet’s success.

Figure 4 visualizes how key academic factors—attendance, test scores, parental engagement, and homework completion—vary across student performance groups (Low, Medium, High). Notably, students in the ‘Low’ category showed 50% lower attendance and 30% less parental involvement, highlighting areas for timely intervention. This highlights the model’s interpretability and usefulness for guiding educational strategies. Figure 5 presents feature importance derived from XGB and RFE, ranking attributes by their predictive power. Each bar reflects how strongly a feature influences performance classification. High-impact features were prioritized, while less relevant ones were excluded to streamline the model. This selection process improves computational efficiency and prediction accuracy, ensuring that only the most informative attributes guide the XSEJNet model’s decisions. Together, these visualizations reinforce the value of data-driven insights for targeted academic support.

Figure 6 shows eight bar graphs, each indicating a distinct characteristic like ‘Gender,’ ‘Nationality,’ ‘Place of birth,’ ‘StageID,’ ‘GradeID’, and Topic. The color scheme clarifies and classifies. The height of the bars represents the percentage distribution of attribute values. The grid simplifies attribute comparison. This figure helps explain dataset trends, student performance analysis, and prediction by revealing attribute distributions and correlations.

To better understand and predict student performance, we compared the proposed XSEJNet model with various other methods, as shown in Table 5. This comparison encompasses both classic machine learning algorithms and more recent, sophisticated models, such as RLCHI (Vimarsha et al., 2024); Enhanced AEO-XGBoost (Cheng, Liu & Jia, 2024), convolution-based deep learning (Conv-DL) (Alshamaila et al., 2024), and DualGNN (Huang & Zeng, 2024). We evaluated the accuracy prediction ability of each model using several key performance metrics. These included accuracy (how often the model was correct), precision (how many of the predicted positives were accurate), recall (how well the model captured all relevant positive cases), and the F1-score, which strikes a balance between precision and recall. We also examined the AUC, which indicates how well a model can distinguish between classes, and the log loss, which measures the confidence of its predictions. Across the board, XSEJNet consistently outperformed all other models, showing at least a 6% improvement in every metric. This highlights its strong predictive power and reliability in real-world educational settings.

The model’s training-test accuracy and loss are shown in Fig. 7. Although there is a clear disparity between the training and assessment validation findings at the outset, the model demonstrates improved learning and training convergence with an increase in the number of epochs. As the difference between training and test performance decreases, it shows that the model can generalize effectively and perform well with unknown input. A clear picture of the model’s ability to adapt to the dataset throughout several epochs and its training dynamics is given by the figure.

The proposed XSEJNet model outperforms previous state-of-the-art approaches with an accuracy of 97.46% on all evaluation measures. Figure 8 displays a comparison of performance indicators and confusion matrix, revealing a close match between expected and real labels. Importantly, the XSEJNet model has much less false positives (FPs) and false negatives (FNs) than previous techniques. This improves accuracy and lowers misclassification, which is crucial in academic decision-making. The confusion matrix in Fig. 8 demonstrates that the suggested technique balances sensitivity and specificity, correctly identifying both successful and at-risk pupils. XSEJNet improves automated educational evaluations by eliminating false alarms and ignored situations. Its excellent accuracy and error control make it a good contender for student performance analytics.

The proposed XSEJNet model was tested for computational efficiency and training duration to determine its applicability and scalability. All tests were conducted on a standard computer with a Core i7 CPU and 16 GB of RAM, ensuring that the findings are replicable and accessible to most academic and institutional settings. Traditional models like SVM (71 s), ResNet (71 s), and more complex methods like RLCHI (63 s), enhanced AEO-XGBoost (56 s), Conv-DL (48 s), and DualGNN (39 s) were all surpass by XSEJNet, which took an average of 23 s to train as shown in Table 6. With a standard deviation of just 2.11 s, XSEJNet proved to be both quick and consistent throughout several runs. With the use of parallel ResNeXt blocks for feature extraction and squeeze-and-excitation attention methods for dimensional focus, the model’s architecture is optimized to reduce computational overhead. These considerations in the architecture make the model fast during training and adaptable to large datasets and real-time academic deployments, making it a powerful and scalable tool for predicting academic achievement.

Table 7 provides a side-by-side view of how different models perform statistically. XSEJNet leads the pack, scoring the highest in metrics like Kendall’s Tau (0.80), Spearman’s Rho (0.84), Pearson correlation (0.93), and chi-square (10.12), along with a very low p-value (0.007) and a strong ANOVA score (8.62). These results demonstrate that XSEJNet effectively captures complex patterns in the data. Due to its innovative combination of ResNeXt blocks, SE attention, and fine-tuned parameters, the model delivers accurate and reliable predictions, making it a solid fit for academic performance forecasting.

Discussion

The objectives mentioned before the simulation section have been achieved. The details are as follows:

Obj-1: factors that affect students’ ability to predict their academic success:

• This study conducted an exploratory analysis of the dataset, considering variables such as cumulative grade point average (CGPA), test results, internal and external assessments, active engagement in extracurricular activities, etc.

• The distribution and frequency of different elements impacting academic performance are shown in visualizations such as those seen in Fig. 8, and others.

• Features like ‘Class,’ ‘Relation,’ ‘Gender,’ ‘NationalITy,’ ‘PlaceofBirth,’ ‘StageID,’ ‘GradeID,’ ‘Topic,’ ‘Semester,’ and others were analyzed to understand their importance and distribution (Fig. 8).

• Insights into student performance attribute groups are provided through visualizations like Fig. 4, which compares features and feature scores.

Obj-2: Study data mining techniques employed to build and verify the models:

• Feature importance is calculated through the XGB technique (Fig. 5).

• The section discusses a feature selection approach based on feature significance analysis to optimize computing resources.

• A comprehensive comparison of multiple models, including the proposed XSEJNet technique, is presented in Table 5.

• To assess and validate the models, performance measures including log loss, reliability, precision, recollection, F1-score, and AUC are utilized.

This work also discusses the model’s effectiveness and efficiency compared to existing models, including a detailed performance comparison (Table 5) and a confusion matrix (Fig. 8). This demonstrates how well the proposed model addresses the study’s objectives.

Key insights: With a high projected accuracy of 97.46% and the ability to overcome challenges like class imbalance and computational complexity, the results show that the XSEJNet model is a powerful tool for educational purposes. Success is evident from the fact that its performance metrics are greater than those of comparable models in the literature.

Comparison to prior work: When comparing prediction accuracy and computational efficiency, the XSEJNet model was superior to both SVM and XGB-TPOT. By providing a method for forecasting student performance that is both more scalable and easier to understand, our findings add to the current body of work.

Implications: The findings of this research have important practical implications for the improvement of educational systems via data-driven solutions. In order to maximize the use of resources and reliably identify students at risk, the technique equips administrators and instructors with useful tools. All of this points to the promise of using machine learning to decisions made in the classroom.

Societal impact and future implications

Educational institutions increasingly seek ways to identify learning gaps early and respond with effective strategies. The XSEJNet framework helps address this need by offering a reliable approach to monitoring student performance and providing timely support. Whether through mentoring, extra lessons, or customised feedback, the model enables schools and universities to take proactive steps in supporting students before issues escalate. One of its key advantages is its adaptability. With a design that works well even in low-resource settings, XSEJNet can be implemented across diverse educational environments, including urban and rural settings, as well as those with large or small student populations. It also integrates seamlessly into digital platforms, enabling students and instructors to receive real-time performance insights.

Beyond academic tracking, XSEJNet plays a significant role in promoting educational fairness. By identifying often overlooked or underserved students, the framework ensures that support systems are directed where needed. This fosters a more inclusive atmosphere, where teaching is tailored, and resources are thoughtfully distributed. In the broader context, XSEJNet contributes to long-term goals, including increasing graduation rates, improving learning outcomes, and creating equitable access to education. Aligned with global educational priorities, such as the UN’s Sustainable Development Goal 4, it provides policymakers and educators with a scalable solution for developing strategies that promote lasting academic and social progress.

Conclusion

The goal of this research was to provide a new machine-learning framework that may help with student performance forecasting and the fight for better educational opportunities. We outperformed previous models by achieving a high degree of accuracy (98.16%) with the integration of enhanced approaches. Enhancing student outcomes, especially in providing a good education, is emphasized by this study as being significantly aided by early intervention and tailored assistance. The knowledge acquired from this investigation enables educators to make informed decisions and implement effective strategies, fostering an environment of academic excellence. The proposed method excels in statistical analysis and computational complexity, providing educational institutions with a pragmatic solution to ensure quality education for all students. Its comprehensive implications encompass academia, universities, and government organizations, providing a potent tool for optimizing pedagogical approaches, crafting customized curricula, and supporting students in need.

The model’s adaptability to various contexts and datasets will be explored in future research. Furthermore, an exploration of ensemble techniques and alternative optimization algorithms has the potential to further enhance its predictive capabilities. Our analysis uses 8,009 records and 16 characteristics, although it has limitations. The dataset’s slight class imbalance may affect prediction accuracy despite mitigating attempts. Some characteristics, including “Parent School Satisfaction,” may not be applicable across all school systems. Future research must demonstrate scalability to larger datasets and real-time deployment in educational contexts. Addressing these constraints will enhance the model’s flexibility and practicality in various academic settings.

Supplemental Information