Predicting academic performance for students’ university: case study from Saint Cloud State University

Setting training and testing data

After pre-processing, the dataset contains 17 features and 71,277 records for all the investigated colleges and departments. The dataset has at least four records for each single student to utilize the previous terms to predict term GPA better. To enable us to split the data by considering all students who have successfully completed four terms and more. At the college-level, we consider all the students from various departments in the targeted college. For each student, the last two records are split as a testing set and the prior records as a training set. For instance, if the student has seven records, the first five records are considered as training, and the last two records as testing. Besides, at the department-level, all the records for all departments (except the target department) within each college as a training set, and the records for the particular target department as a testing set. After that, the training set will be used to train the LSTM model for predictions, which produces a trained LSTM model that will be used to predict the term GPA in the testing set.

LSTM model

The impact of the LSTM model has been widely used in natural language modeling, speech, text, machine translation, and other domains. Various studies used the LSTM to accurately predict students’ future performance based on historical data since the LSTM can uncover the dependencies and insights that help optimize learning strategies (Alanya-Beltran, 2024; Wan et al., 2023b; Wang, 2024). LSTM is particularly used to capture the temporal dependencies in the academic paths of students from the previous academic terms to the current term.

The proposed approach uses LSTM as an prediction model. The LSTM model facilitates the preservation of information flow (Febrian et al., 2023), and mitigate challenges of long term dependencies. In particular, the LSTM model incorporates a series of recurrent module which is considered as a specialized variant of recurrent neural network (RNN). The LSTM model propagates information through connection links that is inherently dependent on temporal steps. It is technically implemented through a recurrent cell structure, where individual cells are interconnected across time steps, as illustrated in Fig. 2. This mechanism handles a stable cell state and maintain the reliability preservation of gradient based information, thereby mitigating the risk of information degradation over time.

LSTM is the best-fit model for this study because the dataset consists of short sequential academic records (2–6 terms per student) over 8 years, making temporal modeling essential (Prottasha et al., 2022). LSTM models are known for handling short- to medium-range dependencies as indicated in the studies (Yousafzai et al., 2021; Hochreiter, 1997), and they have been widely used in educational contexts for modeling student progress over time (Piech et al., 2015). With over 71,000 records from 29,455 students, the data size is sufficient for LSTM but not ideal for more complex models like Transformers. Transformer-based models are powerful, however because of their complexity and extensive number of parameters, they typically need a lot more data to train efficiently (Wu et al., 2019, 2021; Baz, 2024). In contrast, LSTMs (Shakor & Khaleel, 2025; Ghazvini, Sharef & Sidi, 2024) are more data-efficient and simpler to train, and it has sequential learning power which improves feature extraction. Thus, it has superior ability to optimize the evolving academic patterns of students.

As displayed in Fig. 2, the LSTM model is attributed by two main parts: the hidden state $H(t)$, which is dynamically updated over time, and the cell state $C(t)$, which captures long-term memory. The cell state $C(t)$ is modulated along the top horizontal line of the LSTM cell, where information is added or removed during a series of gating mechanism. The behavior of each gating element is controlled by learning parameters or weights, which optimized during the training phase. The LSTM cell consists of three essential gates: forget gate, input gate, and output gate. The forget gate $F(t)$ regulates the influence of the input $X(t)$ and the previous hidden state $H(t-1)$ on the cell state $C(t)$. This gate decides whether to retain or discard information from $X(t)$, with $H(t-1)$ being checked on the binary output of a sigmoid activation function. The input gate, consisting $I(t)$ and $I(t)$, controls the degree to which the input value is integrated into the cell state $C(t)$. The output gate $O(t)$ is responsible for generating predictions at each time step by determining the final output by considering the current cell state $C(t)$. The operations associated with these gates are mathematically formalized in Eqs. (1)–(6).

Furthermore, the LSTM model uses four fundamental functions: the sigmoid ( $\sigma$), which produces gate activation values; hyperbolic tangent function ( $tanh$), which scales values to a normalized range, and multiplication ( $\times$) and summation (+), which are used to combine and transform data. These functions collectively enable efficient weight adjustments during the back propagation process. Where W and B represent the weight matrix and bias vector. The sigmoid function is expressed as $sigma(cdot)$, and the hyperbolic tangent function is denoted by $tanh(cdot)$ (Jiang et al., 2021). The weights associated with the forget gate, input gate, and output gate are represented as $W_f$, $W_i$, and $W_o$, respectively. The LSTM layers capture sequential patterns while handling the issue of vanishing gradients (Sangiorgio & Dercole, 2020). The weights across the entire network, LSTM layers throughout back propagation and gradient descent, are iteratively updated to optimize model performance.

(1) $$F(t)=\sigma (W_f[H(t-1),X(t)]+B_f)$$ (2) $$I(t)=\sigma (W_i[H(t-1),X(t)]+B_i)$$ (3) $$O(t)=\sigma (W_o[H(t-1),X(t)]+B_o)$$ (4) $$\underset{\_}{I}(t)=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}(W_i[H(t-1),X(t)]+B_i)$$ (5) $$C(t)=F(t)\cdot C(t-1)+I(t)\cdot \underset{\_}{I}(t)$$ (6) $$H(t)=O(t)\cdot \tanh (c(t))(\tanh (x)=\frac{\sinh }{cosh}=\frac{e^x-e^{-1}}{e^x+e^{-1}}).$$

The architecture of the LSTM consists of two layers, each containing 64 units, followed by dense output layer. The number of units in each layer defines the dimensionality of the output space. The first layer is designed to capture short term dependencies between elements within the sequence, while the second layer identifies long term temporal dependencies from extended sequences. Once the LSTM layers have extracted temporal features from the input sequences, a dense layer is employed to map these features into the output space for final prediction. After that, during the compilation step, the optimizer and evaluation metrics are specified. The proposed approach utilizes Adam optimizer (Arya & Hanumat Sastry, 2022), which improves the model performance by adaptively adjusting the learning rate throughout the training phase. Finally, the predicted values for college-level and department-level are compared to the actual values during the evaluation step to assess the accuracy of the model.

Evaluation metrics

This study deploys the evaluation metrics to assess the model’s performance (Mansour, Obeidat & Hawashin, 2023; Kanan et al., 2023, 2022). The metrics of mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), R², and mean absolutel percentage error (MAPE) expressed in Eqs. (7)–(11).

• 1.

  MAE: is a measurement that calculates the average size of the errors in a group of predictions, regardless of their direction.

  (7) $$\mathrm{M}\mathrm{A}\mathrm{E}=\sum _{i=1}^D|x_i-y_i|.$$

• 2.

  MSE: it calculated as the difference between predicted ( $\hat{y}$) and actual (y) values.

  (8) $$\mathrm{M}\mathrm{S}\mathrm{E}(y,\hat{y})=\frac{{\sum }_{i=0}^{N-1}{(y_i-{\hat{y}}_i)}^2}{N}.$$

• 3.

  RMSE is a frequently used metric for assessing the precision of a model in regression analysis. RMSE is the square root of the average of the squared differences between predicted and actual values.

  (9) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}(y,\hat{y})=\sqrt{\frac{{\sum }_{i=0}^{N-1}{(y_i-{\hat{y}}_i)}^2}{N}}.$$

• 4.

  R-squared ( $R^2$) assesses the accuracy of a regression model by showing how much variation in the outcome variable can be explained by the explanatory variables.

  (10) $$R^2(y,\hat{y})=1-\frac{{\sum }_{i=1}^N{(y_i-{\hat{y}}_i)}^2}{{\sum }_{i=1}^N{(y_i-\overline{y})}^2}.$$

• 5.

  Mean Absolute Percentage Error (MAPE): measures the accuracy of the forecasting model, it calculates the average percentage difference between predicted and actual values, where a lower MAPE signifies higher prediction accuracy.

  (11) $$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}(y,\hat{y})=\frac{100\mathrm{\%}}{\mathrm{N}}\sum _{\mathrm{i}=0}^{\mathrm{N}-1}\frac{|y_i-{\hat{y}}_i|}{|y_i|}.$$

Dataset description

In this experiment, we collected a real dataset of undergraduate students from a public 4-year institution, namely Saint Cloud State University, for a time of 8 years (2016–2024). This data set contains 108,336 records of anonymous students with 20 characteristics for 29,455 students. Each record represents a single term for a particular student. The data set contains various features related to student demographics, academic performance, and enrollment details. Table 1 shows the description for all features. The data set contains five academic colleges and 46 academic departments. Table 2 shows statistics about these colleges and departments.

Experimental setting

The proposed method is developed using the Python programming language with the aid of tensorflow, keras, sklearn, pandas, and numpy packages, on an Intel(R) Core i7 CPU operating at 2.00 GHz with 16 GB of RAM. Table 3 shows the parameter setting in this research. The LSTM model consists of two layers with 64 units each, followed by a dense output layer. It was trained for 100 epochs with a batch size of 64, using the Adam optimizer with learning rate = 0.001 and ReLU activation. The MSE loss function was used, and 20% of the training data was set aside for validation. To prevent overfitting and improve convergence, a ReduceLROnPlateau callback was applied.

In addition, the experimental settings used various parameters. First, a sequence length of 2 was chosen to allow the LSTM model to capture short-term temporal patterns between consecutive academic terms (Brownlee, 2017; Graves & Graves, 2012). This length balances the need for time-series learning with data availability across all students. Also, using a length of 2 ensures consistent, reliable inputs and supports accurate early prediction without overfitting. Second, the label encoding was used instead of one-hot encoding to reduce feature dimensionality and avoid sparsity, which can negatively impact LSTM performance. One-hot encoding would have significantly increased the input size due to many categorical variables. Label encoding, combined with normalization, offers a more efficient representation while maintaining model stability and performance in a sequence-based context. Third, MinMaxScaler was utilized because it scales data to [0, 1], which aligns well with LSTM’s use of sigmoid and tanh activation functions, enhancing training stability and convergence. MinMaxScaler helps prevent vanishing gradients. It was also preferred over RobustScaler since the dataset was pre-cleaned with minimal outliers. This choice improves model performance by ensuring inputs remain within an optimal range for time-series learning.

To further support the rationale behind these architectural settings, the experimental settings were selected based on theoretical considerations and insights from recent literature in time series modeling (Waqas, Humphries & Hlaing, 2024; Alanya-Beltran, 2024; Wan et al., 2023a; Landi et al., 2021; Waqas et al., 2024; Baz, 2023). The two layers design balances model complexity and generalization, and it allows to capture both short-term and long-term academic data without overfitting. The use of 64 units per layer aligns with established practices in sequential data modeling (Jiang et al., 2021), this dimensionality supports sufficient learning capacity while preserving computational efficiency. ReLU was chosen as the activation function for its ability to mitigate vanishing gradient, and Adam optimizer with learning rate of 0.001 is widely used as robust for adaptive learning in LSTM context (Arya & Hanumat Sastry, 2022). These hyperparameters were determined through empirical tuning to optimize prediction accuracy and training stability.

Features importance

Feature importance involves calculating the score for all input variables to identify significant features through prediction. The higher feature score, the larger impact on the prediction model. Under the permutation method (Huang, Lu & Xu, 2016; Altmann et al., 2010), the feature importance is calculated by noticing the increase or decrease in error when we permute the values of a feature. If permuting the values causes a huge change in the error, it means the feature is important for our model. Figure 3 shows the correlation coefficient between different features and term GPA.

The feature importance results show that academic variables (TermCreditsNotEarned, LocalCreditsEarned, and TermEnrolledCredits) had the strongest impact on term GPA prediction, while demographic and external factors (Gender, Residency) had lower predictive weight. This indicates that academic engagement plays a more direct role in student performance than static personal attributes. Such features, TermCreditsNotEarned, LocalCreditsEarned, and TermEnrolledCredits, directly measure students’ academic engagement and performance within a given term and dynamically linked with with academic behavior patterns that the LSTM model is designed to capture. In contrast, Gender and Residency are static demographic attributes that do not change over the time or reflect actual academic activity as they only indirectly related to academic performance, which limits their predictive power in a time-series model like LSTM. Including these demographic features still enhances transparency by allowing the model to assess their influence fairly, and their lower importance supports the model’s focus on actionable academic indicators, reinforcing both validity and explanation of the findings.

Features as (CourseModality and TransferStudent) were included as input features to enhance model robustness. Course modality influences academic performances due to differences in instructional delivery, students’ participation, and access to resources. Including this feature enable the model to adjust to various learning environments. Likewise, Transfer students are a critical demographic factor, as they often lack to have a consistent academic history and encounter transnational challenges as transfer credits issues, which can impact term GPA. To handle this, the model uses available academic records while normalization and sequence filtering ensures that only students with at least four academic records are included. These features improve the generalizability of the prediction model through capturing diverse academic pathways and institutional experiences.

As term GPA relies on number of credit hours, the TermCreditsNotEarned represents a strong positive correlation with GPA prediction. The reason is when students have more remaining credit hours tends to provide the prediction model with more academic records which leads to obtain a higher prediction performance. Similarly, The TermLocalCreditsEarned shows when students earned more local credits, the LSTM model can perform higher prediction. The TermLocalCreditsEarned impact the prediction model. In Fig. 4, it shows students with low term GPA in College of Education and Learning Design (CoELD) has difference between actual and predicted values when it compared to College of Science and Engineering (CoSE). The CoELD adapts performance driven assessment while the CoSE is content driven assessment. Courses in the CoELD have training pedagogy and constructed based on different teaching and assessment styles. Variance in assessment makes effect on prediction performance. Furthermore, the TermEnrolledCredits shows the higher enrolled credits in a particular academic term which tends to have higher prediction results.

Switching course modality influences outcomes and assessment strategies, during COVID-19 pandemic, SCSU developed different course modality, in-person, on-line, and hybrid. Some students prefer to take online modality rather than in-person, this may causes changes in term GPA. Normally, transfer students are moving from a 2-year community college to new structure of 4-year university. The transfer students may show low term GPA in the first and second terms because the learning environment is different. As the proposed model mainly focuses on the prediction for third year and fourth year of study, transfer students do not have previous term GPA data as regular students in SCSU. As result, this would impact the prediction accuracy.

It is worth noting that there are certain influential factors such as course difficulty, class size, instructor effects, socioeconomic variables were excluded in this study. The primary reason is the unavailability of such features due to the university policy and regulations within the investigated dataset. Further, integrating behavioral or contextual features often requires cross-system access, which raises privacy and consistency challenges.

College-level results

The results represent that students majoring in different colleges have different term GPA. Standing on the educational perspectives, this study provides a justification of why the results shows slightly different prediction accuracy.

Based on RMSE value, Table 4 shows the highest, medium, and the lowest predication accuracy in colleges are Science and Engineering, Herberger Business, and Education and Learning Design, respectively. The variations in prediction accuracy can be explained by the differences in the courses offered and assessment styles by these colleges (Sin & Soares, 2020). The College of Science and Engineering, traditionally provides content-driven courses, where the assessments are based on technical measurable outputs as conducting labs, coding, experiments, and design solutions. The Herberger Business College offers completion-driven course, where assessments are constructed based on applying knowledge to real world applications, scenarios, code (Al-Ahmad et al., 2023), projects, and case studies. On other hand, the College of Education and Learning Design has communication-driven courses, where the assessments are mostly subjective, construed on respecting multiple teaching styles and interpersonal skills between teachers and different learners as adults, children, or special needed students. For example, in the classroom management course, instructors teach social emotional learners and regular learners. Consequently, assessment of such course is different from lab or experiment test as conducted in the college of Science and Engineering. These factors introduces inconsistencies that challenges the prediction model. Obviously, subjective assessment is harder than objective assessment and this would make a variance of the prediction performance of term GPA.

Figure 4 illustrates the scatter plots of predicted vs. actual term GPA values in all investigated colleges over the test set, showing the achieved RMSE values at the top of each sub-figure. The differences are reported as 5%, 2.8%, 4.8%, 2.6%, and 3.2%, for College of Education and Learning Design, College of Health and Wellness Professions, College of Liberal Arts, College of Science and Engineering, and Herberger Business College, respectively. College of Science and Engineering highlights most accurate prediction and College of Education reflects least accurate performance. While most colleges exhibit relatively small differences, which indicate strong alignment between predicted and actual term GPA values, slight variability exist depending on characteristics of each college.

In addition, Fig. 5 represents the model’s learning curves with the loss values compared to training epochs for both the training and validation sets, indicating a normal training behavior without overfitting in the proposed model. It is clearly shown that the validation loss roughly corresponds to the training loss for the College of Education and Learning, College of Science and Engineering, and Herberger Business College, with no notable divergence, indicating that the model is not overfitting and can generalize well to new data. On the other hand, and for the College of Health and Wellness Professions and the College of Liberal Arts, there is a small divergence during the early epochs, when the validation loss marginally increases while the training loss reduces. This early gap is likely a result of an infrequent overfitting phase in which the model attempts to memorize the training data before completely capturing generalizable patterns.

The discrepancy in prediction accuracy across different colleges stems from differences in evaluation styles and curriculum structure. For example, College of Science and Engineering tends to use objective, quantifiable assessment such as exams, lab reports, and problem solving tasks, which result in more consistent term GPA patterns and higher model accuracy. In contrast, College of Education and Learning Design rely on subjective based evaluation, such as classroom engagement and teaching simulations, which lead to have more variability and make it harder to predict term GPA. These differences impact how well LSTM model can learn and generalize patterns.

Department-level results

Evaluation of students depends on certain factors as time management (Wolters & Brady, 2021), students motivation (Spurlock, 2023), students roles in homogeneous or heterogeneous groups (Briggs, 2020). Also, different university may have different academic structures (Delbanco, 2023) as courses, program maps, or majors. If the structure is harder, the low-expectation students encounter some challenges that increases possibilities of failure or having low term GPA. This study aims to explore why there are prediction variations among departments in the same college. These differences can be attributed to the nature of courses and assessment methods within each department.

As shown in Table 5, the highest, medium, and lowest prediction accuracy of departments are shown in Information Media, Special Education, and Child and Family Studies, accordingly. The courses in Information Media concentrating on how to use e-Learning and online media tools to deliver content knowledge to K-12 students or special-needed students. The assessments are structured on integrating technology to improve the learning process. The technology and relatively objective evaluations contributes to higher prediction accuracy of term GPA. The courses in Special Education department are designed to address different types of special needed students. Even though rich theories are taught in special education courses, but the assessments are still based on how much improvements achieved by special education students compared to the normal students. The variability in learning among special needed students indicates to find moderate prediction performance. The courses from the Child and Family Studies department provide students to learn how to develop the creativity of children and family metrics. Comparing to the previous two departments, it is reasonable to observe that the prediction performance is slightly lower in this department as it conducts qualitative assessment, which results in challenges to have consistent evaluations.

Referring to College of Health and Wellness results displayed in Table 6, the highest, medium, and lowest prediction accuracy of departments are observed in Medical Laboratory Science, Nursing, and Kinesiology, respectively. The courses from Medical Laboratory Science department emphasize applying scientific process to evaluate students’ understanding. Assessment in this department are more objective and rely on standard metrics such as labs procedures, diagnostics accuracy, and medical policies. Such assessments reflect that this department achieves the highest prediction. In the Nursing department, courses focuses more on behavioral and interpersonal aspects, such as understanding human feelings and interactions. Assessments are designed to respect outcomes related to hospital procedures and treatment with patients. The combination of behavioral evaluation and procedural metrics results in medium prediction performance. Courses in the Kinesiology departments incorporate a broader scope of content knowledge, societal awareness, and human characteristics. Assessments in this department often include multiple dimensions of physical environment, human body, and psychological components. The interdisciplinary nature of these courses combined with subjective evaluation, leads to the lowest prediction.

With regard to College of Liberal Arts, as indicated in Table 7, the highest, medium, and lowest prediction accuracy of departments are occurred in Language and Culture, Theater and Film Studies, Anthropology, correspondingly. Courses in the Language and Culture department focus on teaching various languages and promoting cultural understanding (Altarriba & Basnight-Brown, 2022), language reflects culture; as students gain proficiency in language, they develop a deeper understanding of the related culture. This cultural intuition improves their ability to perform better in subsequent terms. Assessments in this departments include mix of objective and subjective methods that make balance grading system, this would help the proposed model to achieve higher accuracy as students improve their linguistic and cultural skills over the time. In the Theater and Film Studies department, courses mainly helps students develop their creative ability in theater and film production. Assessment of such courses are basically subjective metrics. Such subjectivity often leads to variability in evaluations, as students and instructors may have different perspectives and artistic criteria, resulting in moderate prediction. Anthropology courses aim to explore theories of explaining the past and future of human culture and activities. Assessment methods in such discipline are harder to precisely reflect student’s outcomes, which shows lower prediction performance.

In the College of Science and Engineering as represented in Table 8, the prediction accuracy varies across departments. The Physics and Astronomy department reveals the highest prediction values, courses are reflecting strongly consistent and well defined structure. Assessments are mainly objective and they based on theoretical understanding and analytical problem solving skills, leading to better prediction performance. The Chemistry and Biochemistry department demonstrates moderate prediction performance, evaluations are combination of objective and subjective, such as lab experiments and applied instructional methods. Biological Science department shows the lowest prediction value, assessments are primarily subjective and depend on qualitative clinical metrics. Courses are designed as preliminary or prerequisites for medical field as nursing and medical labs, which impacts overall prediction reliability.

The results of Herberger Business College, as given in Table 9, the highest prediction accuracy is observed in the Information Systems department, followed by moderate accuracy in the Management and Entrepreneurship department, and the lowest in Accounting department. Courses in Information Systems department emphasize practical knowledge, coding skills, and software development. The technical based assessment contributes to higher accuracy, as outcomes are more measurable and consistent. Courses in Management and Entrepreneurship are business driven and developed around abstract business concepts as leadership, innovation, and strategic management. The conceptual based assessments presents moderate prediction. The accounting department focuses on content intensive courses that include advanced statistical methods, real world applications, and financial policies. The complexity of courses introduces variance between the predicted and actual results.

At the department level, discrepancies are often due to diversity of course content, heterogeneity of student populations, and subjectivity in grading. Departments like Medical Laboratory Science follows structured course content, consistent learning outcomes, and used standardized evaluations, this enable higher prediction accuracy. Conversely, departments such as Child and Family Studies involve qualitative assessment that are less predictable, resulting in lower model performance. In addition, departments with diverse student backgrounds or interdisciplinary content introduce greater variation in term GPA outcomes, and making prediction harder in these academic contexts.

The plots shown in Fig. 6 through Fig. 10, illustrate the predicted vs. actual term GPA values for highest and lowest RMSE values of departments in each single college: College of Education and Learning Design, College of Health and Wellness Professions, College of Liberal Arts, College of Science and Engineering, and Herberger Business College. College of Education and Learning Design, the highest difference is reported in Child and Family Studies department is 14.6%, whereas the lowest difference is indicated as 3.1% in the department of Information Media. Similarly, in College of Liberal Arts, our findings indicate that there is difference between the actual and the predicted term GPA as 9.6% in Anthropology, and 1.4% in Languages and Cultures department as given in Fig. 8. On the other hand, there is slight differences among departments in Herberger Business College. The findings reveal homogeneity in prediction accuracy within the Herberger Business College, while greater variability is observed in the College of Education and Learning Design.

In summary, subjective evaluation provides variability and less quantifiable grading patterns which make lower prediction whereas objective evaluation introduces more quantifiable consistent data that improve the prediction accuracy. The proposed model will help advisors to observe the academic performance of their advisees through predicting the low-expectation students with failed or low term GPA. Consequently, this would be an effective strategy to avoid unexpected term GPA decline and improve the academic success.

Validation of the proposed model

To validate our proposed approach, we compare our findings with five traditional ML classifiers, two deep learning models, including CNN and RNN, by using the same dataset, as indicated in Table 10, and the related work as presented in Table 11. The ML classifiers are linear regression (LR), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), and support vector regressor (SVR). Among deep learning models, the LSTM had the lowest RMSE of 0.0108, indicating higher prediction accuracy. In contrast, the CNN and RNN models had RMSEs of 0.1479 and 0.1925, respectively. These findings reveal that the LSTM outperforms those two models in terms of capturing temporal relationships and complicated patterns in students academic records. LSTM is designed to detect both short-term and long-term dependencies in sequential data, making it ideal for predicting term GPA across multiple academic semesters. Unlike RNNs, which struggle with vanishing gradients, LSTMs use memory cells and gating mechanisms to preserve all relevant previous information in an effective fashion. CNNs are very efficient for spatial data, and less suited for time-series prediction. Also, compared with the related approaches, our proposed model achieves the highest prediction performance in all reported evaluation metrics. For example, the findings in the study (Prabowo et al., 2021) introduced 0.34 and 0.414 for MAE and MSE. On the other hand, our results reveal 0.0059 and 0.001 for the same metrics.

Obviously, prediction results vary across universities due to differences in academic structure, assessment methods, students demographic, data availability, and how well prediction models align with institutional contexts. Universities use varying grading styles, some rely on objective assessments like exams and labs, while others use subjective evaluations such as projects (Al-Ahmad et al., 2022) and classroom engagement, leading to inconsistencies in GPA patterns. Additionally, factors like transfer status, course difficulty, and the absence of behavioral data influence model accuracy. As a result, even with similar predictive models, performance outcomes differ significantly depending on the educational environment and data characteristics of each institution.

The differences in performance of the proposed approach comparing to other universities stems from institutional, academic, and methodological variation. First, at SCSU, the dataset consists of 8 years and includes over 71,000 records from 29,455 students across five colleges and 46 departments. This provides diverse and comprehensive academic context. This contrasts with most of comparative studies that used small or homogeneous dataset that focused on one or two colleges. Second, the proposed model predicts term GPA at both of college and department levels, and detects deeper variation within the university that other related models may overlook. Third, the proposed model captures only academic and demographic features due to institutional constraints, so the behavioral and non-cognitive features which have been used by the previous related approaches may influence the prediction performance. Such reasons clarify why its performance may differ from the previous studies that conducted in more uniform academic environments.