This section outlines the methodology adopted for our study. Ethical Approvals and Data Source: We first describe the dataset and ethical approvals, followed by data preprocessing, the theoretical framework, the EduTransNet architecture, and the mathematical formulations underpinning the model. The dataset used in this study was collected from three Pakistani universities—Islamia University of Bahawalpur (n = 1023), Bahauddin Zakariya University Multan (n = 987), and Khwaja Fareed University of Engineering and Information Technology Rahim Yar Khan (n = 837)—comprising 2847 student records collected between September 2022 and May 2023. Data collection was conducted under Institutional Review Board (IRB) approval from each participating institution (IRB Protocol Numbers: IUB-2022-EDU-031, BZU-2022-RES-047, KFUEIT-2022-ETH-019). All participants provided written informed consent, and all data were anonymised prior to analysis.
The dataset includes student demographics (gender, age, ethnicity), academic performance metrics (midterm grade, final grade, attendance, study hours), engagement data (engagement score), and ethical perception variables measured on 5-point Likert scales (privacy concerns, algorithmic bias awareness, fairness perception, transparency expectations). The target variable is the transparency score, a composite metric (0–100) described in the Introduction. The data were split as follows: 85% (2,420 records) for model development (further divided into training and validation sets using 10-fold stratified cross-validation) and 15% (427 records) as a completely held-out test set reserved exclusively for final performance evaluation, never used during model selection or hyperparameter tuning. The methodology addresses two core optimisation problems. The first concerns optimising data privacy in AI-based educational systems.
$$L\left({P}_{d}\right)=\sum_{i=1}^{n}\frac{1}{{P}_{d}}log\frac{1}{{P}_{d}}+\frac{1}{1-{P}_{d}}log\frac{1}{1-{P}_{d}}$$
(1)
where n is the size of the dataset.
The privacy level variable Pd is operationalised as a continuous value in the range [0, 1], where 0 represents no privacy protection (full data exposure), and 1 represents maximum privacy protection (complete anonymisation). In practice, Pd = 0.7 was selected based on pilot testing, representing a configuration where 70% of identifying information is obscured through k-anonymity (k = 5) and differential privacy (ε = 1.0) mechanisms applied during data preprocessing.
The objective is to maximise the utility function \(U(Pd,D,M)\) while minimising the privacy loss:
$$\frac{{{\text{max}}}}{{{\text{P}}_{{\text{d}}} }}{\text{U}}\left( {{\text{P}}_{{\text{d}}} ,{\text{D}},{\text{M}}} \right) – \lambda \cdot {\text{L}}\left( {{\text{P}}_{{\text{d}}} } \right)$$
(2)
where λ is a trade-off parameter.
\(PdPrivacy\) level variable. M Machine learning model \(L\left(Pd\right)\) Privacy loss function \(U(Pd,D,M)\) Utility function λ Trade-off parameter.
The privacy level Pd is the variable under consideration, representing the extent of privacy protection in the AI-based educational system. The privacy loss function \(L\left(Pd\right)\) quantifies the privacy compromise associated with a specific level. The objective is to find the optimal privacy level (Pd) that maximises the utility \(\left(U\left(Pd,D,M\right)\right)\) of the system while considering the trade-off parameter (λ) to balance utility and privacy loss.
The second problem concerns ensuring algorithmic transparency in AI-driven educational decision-making. Given an algorithm A and an educational environment E, the goal is to find transparency metrics T that maximise utility while minimising transparency loss.
Let \(T\) be the transparency metrics variable, A represent the algorithm, and E denote the educational environment.
The transparency loss function L(T) is defined as:
$$L\left(T\right)=\frac{1}{2}\sum_{j=1}^{m}\text{}\left|{T}_{j}-\text{M}\text{e}\text{a}\text{n}\left(T\right)\right|$$
(3)
where m is the number of transparency metrics.
The objective is to maximise the utility function U(T, A,E) while minimising the transparency loss:
$$\mathop {\max }\limits_{T} U(T,A,E) – \lambda \cdot L\left( T \right)$$
(4)
where \(\lambda\) is a trade-off parameter.
$$T\quad {\text{Transparencymetricsvariable}}$$
$$A\quad {\text{Algorithm}}$$
$$E\quad {\text{Educationalenvironment}}$$
$$L\left( T \right)\quad {\text{Transparencylossfunction}}$$
$$U(T,A,E)\quad {\text{Utilityfunction}}$$
The optimisation frameworks presented in Eqs. (1)-(4) establish the theoretical foundation for EduTransNet’s design philosophy. Specifically, the privacy-utility trade-off (Eqs. 1–2) informs the model’s data-handling procedures, where the hyperparameter λ corresponds to the regularisation coefficient in our implementation, balancing predictive performance with privacy preservation during training. Similarly, the transparency optimisation framework (Eqs. 3–4) directly translates to EduTransNet’s architecture: the transparency metrics T correspond to the interpretability scores generated by the final sigmoid layer, while the utility function U(T, A,E) is operationalised through the combined loss function that optimises both prediction accuracy and transparency. This theoretical grounding ensures that EduTransNet does not merely predict transparency scores but embodies transparency principles in its architectural design.
The transparency metrics \(T\) represent the quantifiable aspects of algorithmic transparency in educational decision-making. The transparency loss function \(L\left(T\right)\) calculates the deviation of each metric from the mean, indicating the extent of transparency compromise. The objective is to find the optimal set of transparency metrics (\(T\)) that maximises the utility (\(U(T,A,E)\)) of the system while considering the trade-off parameter (\(\lambda\)) to balance utility and transparency loss.
To ensure that the methodology is thoroughly explained, we first detail the dataset used for the study. The dataset comprises educational data sourced including features e.g., student performance metrics, demographic information, etc.]. Before feeding the data into the model, pre-processing steps were performed to clean and prepare it. These steps included handling missing values, normalising feature scales, and splitting the dataset into training and testing sets. Categorical features were encoded as needed, and continuous features were standardised to ensure consistency across the dataset. The core of the methodology is built around the EduTransNet architecture, a novel deep learning model designed explicitly for predicting educational data. The architecture employs multiple layers, which allow the model to capture complex relationships in the data.
Additionally, the model includes mechanisms to ensure fairness and transparency in predictions, a crucial aspect given the ethical implications in educational settings. For the training process, the hyperparameters were carefully chosen after an initial tuning phase. The model was trained with a learning rate as the optimisation algorithm. Other key hyperparameters include the batch size, the number of epochs, and regularisation techniques such as dropout or L2 regularisation to prevent overfitting. During training, the model’s performance was monitored via cross-validation to ensure robustness. The evaluation metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared values, were used to gauge the model’s accuracy and generalizability. Furthermore, early stopping was implemented to prevent overfitting: training was halted if validation performance stopped improving after a specified number of epochs.
Dataset description
The dataset used in this study is privately held and was collected under Institutional Review Board (IRB) approval from each participating university (IRB Protocol Numbers: IUB-2022-EDU-031, BZU-2022-RES-047, KFUEIT-2022-ETH-019). All participants provided written informed consent prior to data collection. The data remain confidential and are not publicly available; however, anonymised subsets may be made available to qualified researchers upon reasonable request and appropriate data-sharing agreements. All methods were performed in accordance with the relevant institutional guidelines and regulations, consistent with the principles of the Declaration of Helsinki and applicable national ethical standards. All participants provided written informed consent prior to participation. The dataset comprises 2,847 student records collected from three Pakistani universities: Islamia University of Bahawalpur (n = 1,023), Bahauddin Zakariya University Multan (n = 987), and Khwaja Fareed University of Engineering and Information Technology Rahim Yar Khan (n = 837). Data collection was conducted between September 2022 and May 2023 following institutional review board (IRB) approval from each participating institution. All data were anonymised prior to analysis, and informed consent was obtained from participants.
The dataset comprises 2,847 student records including the following features (Table 2): Student ID (unique identifier), Gender (categorical: Male/Female/Other), Age (numeric: 18–35 years), Ethnicity (categorical: 5 categories), Course Code (categorical: 47 unique courses), Semester (categorical: 1–8), Attendance (numeric: 0-100%), Midterm Grade (numeric: 0-100), Final Grade (numeric: 0-100), Study Hours (numeric: 0–40 h/week), Engagement Score (numeric: 0-100), Privacy Concerns (numeric: 1–5 Likert scale), Algorithmic Bias Awareness (numeric: 1–5 Likert scale), Fairness Perception (numeric: 1–5 Likert scale), and Transparency Expectations (numeric: 1–5 Likert scale).
The transparency score is a composite metric ranging from 0 to 100 that quantifies the degree to which AI-driven educational decisions are interpretable, explainable, and accountable to stakeholders. It integrates four components: (1) algorithmic explainability—the extent to which the model’s reasoning can be articulated; (2) decision traceability—the ability to audit how inputs led to outputs; (3) stakeholder communication clarity—how effectively predictions can be communicated to non-technical users; and (4) ethical compliance—adherence to established AI ethics guidelines.
Table 2 Feature Description of the Dataset.
The dataset comprises a total of 2847 student records, distributed across the three participating universities: Islamia University of Bahawalpur (n = 1023, 35.9%), Bahauddin Zakariya University Multan (n = 987, 34.7%), and Khwaja Fareed University of Engineering and Information Technology, Rahim Yar Khan (n = 837, 29.4%).
The dataset encompasses students from diverse academic disciplines (engineering, sciences, social sciences, humanities), multiple academic levels (undergraduate and graduate), and varied socioeconomic backgrounds representative of Pakistani higher education. The three participating universities were selected to capture geographical diversity (Punjab province) and institutional variety (one general university, one specialised engineering institution, and one comprehensive research university). The gender distribution (58% male, 42% female) approximates the national higher education enrollment ratios. While the dataset represents Pakistani educational contexts, the underlying transparency-related variables (engagement, fairness perception, privacy concerns) are conceptually transferable to other cultural settings.
Scale definitions for ethical consideration variables
Privacy Concerns (1–5 Likert scale): 1 = No concern, 2 = Slight concern, 3 = Moderate concern, 4 = High concern, 5 = Extreme concern.
Algorithmic Bias Awareness (1–5 Likert scale): 1 = Completely unaware, 2 = Slightly aware, 3 = Moderately aware, 4 = Very aware, 5 = Extremely aware.
Fairness Perception (1–5 Likert scale): 1 = Very unfair, 2 = Somewhat unfair, 3 = Neutral, 4 = Somewhat fair, 5 = Very fair.
Transparency Expectations (1–5 Likert scale): 1 = No expectation, 2 = Low expectation, 3 = Moderate expectation, 4 = High expectation, 5 = Very high expectation.
This table provides a detailed description of the dataset’s features, including student demographics (gender, age, and ethnicity), academic performance metrics (midterm and final grades), and engagement-related data (attendance, study hours, and engagement score). These features collectively offer valuable insights into student behaviour, performance, and engagement within educational settings. Figure 2 shows the Combined Distribution Plots for Numerical Features. Figure 3 shows the Feature Importance.
Fig. 2
Combined distribution plots for numerical features.
Figure 2 reveals the distributional characteristics of key numerical variables. The attendance distribution shows a slight left skew, indicating most students maintain high attendance rates (> 75%). Study hours exhibit a normal distribution centred around 15–20 h per week. Engagement scores exhibit bimodality, suggesting two distinct student populations: highly engaged and moderately engaged learners.
Fig. 3
Figure 3. This visualisation ranks input features by their predictive contribution to the transparency score. Engagement Score emerges as the most influential predictor (importance = 0.23), followed by Transparency Expectations (0.19) and Fairness Perception (0.17). Academic performance metrics (Midterm Grade, Final Grade) are moderately important (0.12–0.14), while demographic features (Age, Gender) contribute minimally, suggesting the model relies primarily on behavioural and attitudinal variables.
The positive correlations between key predictor variables and the transparency score are visually represented in Fig. 4, which illustrates the linear relationships through scatter plots with regression lines.
Fig. 4
Scatter Plots with Regression Lines.
Figure 4 these plots illustrate bivariate relationships between key predictors and the transparency score. The Engagement Score vs. Transparency plot shows a strong positive linear relationship (r = 0.78), validating engagement as a key predictor. The Study Hours vs. Transparency plot reveals a moderate positive association with some heteroscedasticity at higher study hours.
Data pre-processing
Before training the model, the dataset undergoes meticulous preprocessing to ensure its quality and suitability for analysis. The following steps are performed as part of the data preprocessing pipeline:
Handling Missing Values: Missing data points, if any, are carefully imputed using appropriate methods, such as mean, median, or regression imputation. This ensures that the dataset remains complete and representative.
Feature Scaling: Numeric features are scaled to a standardised range, typically 0–1, using techniques such as Min-Max scaling or standardisation. This step helps in mitigating the influence of features with larger magnitudes and ensures uniformity in feature scales.
Categorical Encoding: Categorical variables are encoded as numerical values using techniques such as one-hot or label encoding. This transformation enables the incorporation of categorical features into machine learning models, which typically require numerical inputs.
Outlier Detection and Removal: Outliers, which may adversely affect model performance, are identified and subsequently removed or treated using robust techniques such as trimming or transformation. This step helps improve the model’s robustness and generalizability.
Feature Engineering: Additional features may be derived or engineered from existing ones to capture complex relationships and patterns in the data. This may involve creating interaction terms, polynomial features, or domain-specific transformations tailored to the educational context.
By meticulously conducting these preprocessing steps, we ensure that the dataset is appropriately formatted, cleaned, and enriched, setting the stage for practical model training and evaluation.
Theoretical frameworkMachine learning and predictive modelling theory
At the core of this research is the application of machine learning (ML) and predictive modelling theories. These theories focus on using algorithms to identify patterns in data and make accurate predictions. EduTransNet is built upon this foundation, leveraging deep learning—a subset of machine learning—as its primary approach. Deep learning models, particularly neural networks, are designed to automatically learn complex features from large datasets, enabling EduTransNet to outperform traditional models such as SVR, LR, and RFR.
This theoretical framework justifies the use of data-driven models in education, where patterns in student performance, behaviour, and demographic data can be identified to predict future outcomes. The core principles guiding this work are derived from supervised learning and regression analysis, which have been widely applied in educational data mining to predict variables like student success and course outcomes. This aligns with the broader data science framework in education, where models inform decisions and improve educational outcomes through evidence-based predictions.
Methodological Implication
Machine learning and predictive modelling theory directly informed three critical design decisions in EduTransNet. First, the choice of a deep neural network over simpler models (SVR, LR, RFR) was justified by the need to capture non-linear, multi-dimensional interactions among educational variables—such as the interplay between engagement scores, fairness perceptions, and demographic factors—that linear models cannot adequately represent. Second, the five-layer architecture with progressively decreasing neuron counts (128→64→32→16→8) was designed to implement hierarchical feature abstraction, where early layers capture low-level statistical patterns (e.g., correlations between attendance and grades) and deeper layers synthesise high-level representations (e.g., the composite relationship between ethical perceptions and transparency). Third, the supervised regression framework was selected because transparency scores are continuous variables (0–100), making regression analysis the most theoretically appropriate approach for prediction.
Fairness and ethical AI
A significant part of this research also draws from ethical AI and fairness in machine learning theories. These theories address the growing concern that unchecked AI models can perpetuate or exacerbate societal biases. The ethical AI framework emphasises transparency, accountability, and fairness in algorithmic decision-making, all of which are critical in educational contexts where AI-driven predictions can impact a student’s academic trajectory.
EduTransNet is guided by fairness-aware machine learning theory, which suggests that models should be evaluated not only for their accuracy but also for their potential to create or reinforce biases. This informs the methodology, where fairness metrics are incorporated into the model evaluation process, ensuring that the model’s predictions do not disproportionately disadvantage any group based on demographic factors. In this context, algorithmic fairness is a key theoretical underpinning, influencing decisions on data preprocessing, model training, and evaluation, as well as the design of transparency mechanisms that enable human oversight.
Methodological Implication:
Fairness and ethical AI theory translated into three concrete implementation decisions. First, EduTransNet’s loss function was augmented with a demographic parity penalty term: L_total = MSE(y, ŷ) + λ_fair · |mean(ŷ_group1) − mean(ŷ_group2)|, where λ_fair = 0.1 was determined through grid search over [0.01, 0.05, 0.1, 0.2] to balance accuracy and fairness. This ensures that mean predicted transparency scores remain comparable across gender and ethnicity groups. Second, equalised odds verification was implemented as a post-training validation step, checking that true positive and false positive rates differed by less than 5% across protected groups. Third, during data preprocessing, stratified sampling was employed to ensure proportional representation of all demographic groups in both training and validation sets, preventing the model from learning biased representations due to class imbalance.
Educational theory and predictive analytics
In addition to machine learning theories, this research is grounded in educational theory. Predictive modelling in education is closely linked to learning analytics, which focuses on collecting, measuring, and analysing data about learners and their contexts. The goal of learning analytics is to understand and optimise learning experiences and outcomes.
EduTransNet operates within this theoretical framework by predicting student performance, identifying at-risk students, and offering personalised course or career recommendations. These predictive analytics approaches are informed by theories of student engagement, academic achievement, and personalised learning, and interventions are designed based on individual student needs. By applying these theories, the model aims to provide timely insights that can help educators take proactive measures to support student success.
Methodological Implication:
Educational theory and learning analytics directly shaped three aspects of EduTransNet’s design. First, feature selection was guided by established learning analytics research: engagement scores, attendance rates, and study hours were included because they are empirically validated predictors of student success, while ethical perception variables (privacy concerns, fairness perception, algorithmic bias awareness, transparency expectations) were included to capture students’ subjective experiences with AI systems—a dimension largely absent in existing models. Second, the composite transparency score was constructed by integrating four theoretically grounded components (algorithmic explainability, decision traceability, stakeholder communication clarity, ethical compliance) that reflect the multidimensional nature of educational transparency as described in the learning analytics literature. Third, the model’s output was designed as a continuous score (0–100) rather than a categorical classification, enabling nuanced, granular predictions that support personalised educational interventions aligned with individual student needs.
Transparency and interpretability in AI
Another key theoretical component of this research is the concept of transparency and interpretability in AI, which is becoming increasingly important in ethical AI frameworks. This theory advocates AI models that are not only accurate but also interpretable, meaning their decisions can be understood and scrutinised by humans. In educational settings, this is particularly important, as educators and administrators must trust and understand how a model arrived at a particular prediction.
EduTransNet is informed by this theory through its integration of transparency scores, which measure how clearly human users can understand the model’s predictions. This aligns with the broader framework of explainable AI (XAI), which seeks to bridge the gap between powerful, complex models and the need for human interpreters to make informed decisions based on AI predictions. The methodological choices regarding model transparency and ethical considerations are rooted in this framework, ensuring that the model can be trusted and used responsibly in educational environments.
Methodological Implication:
Transparency and interpretability theory informed four specific architectural choices. First, a sigmoid activation function was applied in Hidden Layer 5 to produce bounded outputs in the range [0, 1], which are then scaled to [0, 100] to generate transparency scores that are directly interpretable by non-technical stakeholders such as educators and administrators. Second, the progressive layer reduction (128→64→32→16→8→1) was deliberately designed to create an information bottleneck that forces the network to learn compressed, interpretable representations rather than memorising raw data patterns. Third, feature importance analysis using permutation-based methods was integrated as a post-hoc interpretability mechanism, enabling stakeholders to understand which input variables (e.g., engagement score, fairness perception) most strongly influence predictions. Fourth, the architectural depth was limited to five hidden layers—rather than deeper architectures—to maintain a balance between representational capacity and human interpretability, consistent with explainable AI principles that advocate for the simplest model capable of achieving the task.
Bias-variance tradeoff and model generalisation
From a technical standpoint, the research is also guided by the bias-variance tradeoff theory in machine learning, which states that a model’s performance is influenced by its ability to balance bias (errors from incorrect assumptions in the model) and variance (errors from sensitivity to small fluctuations in the training data). This theory is critical for understanding why EduTransNet, through techniques such as regularisation and cross-validation, outperforms simpler models like linear regression, which often suffer from high bias, and more complex models like random forests, which can suffer from high variance.
This concept informs the methodology, as cross-validation techniques are used to assess model generalisation and ensure that EduTransNet is robust when applied to unseen data. This is particularly important in educational contexts, where the model must generalise across diverse student populations with varying characteristics.
Methodological Implication
The bias-variance tradeoff theory directly determined five regularisation and validation decisions. First, L2 regularisation (λ = 0.0001) was applied to all weight matrices to penalise large weights and reduce model variance, preventing the network from fitting noise in the training data. Second, dropout (rate = 0.3) was applied after each hidden layer to randomly deactivate neurons during training, effectively creating an ensemble of sub-networks that improves generalisation. Third, early stopping with a patience of 20 epochs was implemented to halt training when validation loss ceased improving, preventing the model from over-learning training-specific patterns. Fourth, 10-fold stratified cross-validation was used to obtain robust performance estimates that account for data variability across different student subpopulations. Fifth, a held-out test set comprising 15% of the total data (427 records) was reserved for final model evaluation, providing an unbiased estimate of performance on completely unseen data. These combined strategies ensure that EduTransNet achieves high accuracy while maintaining robust generalisation across diverse student populations.
EduTransNet
The EduTransNet for this study is a deep neural network architecture designed to predict student academic performance based on various input features, as shown in Fig. 5. The architecture comprises several layers, each serving a specific purpose in the learning process.
EduTransNet is a novel deep neural network architecture designed to predict transparency scores in educational settings. This section provides a detailed explanation of the proposed model, including its architecture, activation functions, and equations. Figure 6 shows the architecture of the proposed model.
EduTransNet Architecture
The proposed EduTransNet architecture (illustrated in Fig. 5) consists of an input layer that accepts n features, followed by five hidden layers with progressively refined representations, and an output layer that produces the transparency score. The architectural design is as follows.
Input Layer: Accepts the preprocessed feature vector X = [x₁, x₂, …, xₙ].
Hidden Layer 1: 128 neurons with ReLU activation.
Hidden Layer 2: 64 neurons with Tanh activation.
Hidden Layer 3: 32 neurons with ReLU activation.
Hidden Layer 4: 16 neurons with Leaky ReLU activation (α = 0.01).
Hidden Layer 5: 8 neurons with Sigmoid activation.
Output Layer: Single neuron with linear activation producing the transparency score.
Hyperparameters:
The model was trained with the following hyperparameters: learning rate = 0.001, Adam optimiser, batch size = 32, number of epochs = 200 with early stopping (patience = 20), dropout rate = 0.3 for regularisation, and L2 regularisation coefficient = 0.0001.
Activation Functions:
ReLU (Rectified Linear Unit):
$$ReLU\left(x\right)=max\left(0,x\right)$$
(5)
Tanh (Hyperbolic Tangent):
$$Tanh\left(x\right)=\frac{ex-{e}^{-x}}{ex+{e}^{-x}}$$
(6)
Leaky ReLU:
$$LeakyReLU\left(x\right)=\left\{\begin{array}{c}xifx\ge\\0ifx<0\\otherwise\end{array}\right.$$
(7)
where \(\alpha\) is a small constant.
Sigmoid:
$$Sigmoid\left(x\right)=\frac{1}{1+{e}^{-x}}$$
(8)
The following equations govern the computations within each layer of EduTransNet:
Input Layer:
$$X=\left[x1,x2,\dots,xn\right]$$
(9)
where X is the input feature vector.
Hidden Layer 1:
$$\text{H}\left(1\right)=\text{R}\text{e}\text{L}\text{U}\left(\text{W}\left(1\right)\text{X}+\text{b}\left(1\right)\right)$$
(10)
where:
\(W1\) is the weight matrix for Hidden Layer 1.
\(b1\) is the bias vector for Hidden Layer 1.
Hidden Layer 2:
$$H\left(2\right)=Tanh\left(W\left(2\right)H\left(1\right)+b\left(2\right)\right)$$
(11)
where:
\(W2\) is the weight matrix for Hidden Layer 2.
\(b2\) is the bias vector for Hidden Layer 2.
Hidden Layer 3:
$$H\left(3\right)=ReLU\left(W\left(3\right)H\left(2\right)+b\left(3\right)\right)$$
(12)
where:
\(W3\) is the weight matrix for Hidden Layer 3.
\(b3\) is the bias vector for Hidden Layer 3.
Hidden Layer 4:
$$H\left(4\right)=LeakyReLU\left(W\left(4\right)H\left(3\right)+b\left(4\right)\right)$$
(13)
where:
\(W4\) is the weight matrix for Hidden Layer 4.
\(b4\) is the bias vector for Hidden Layer 4.
Hidden Layer 5:
$$H\left(5\right)=Sigmoid\left(W\left(5\right)H\left(4\right)+b\left(5\right)\right)$$
(14)
where:
\(W5\) is the weight matrix for Hidden Layer 5.
\(b5\) is the bias vector for Hidden Layer 5.
Output Layer:
$${y}^{=W\left(O\right)H\left(5\right)}+b\left(O\right)$$
(15)
where:
y^ is the predicted transparency score.
\(W\left(O\right)\) is the weight matrix for the output layer.
\(b\left(O\right)\) is the bias vector for the output layer.
Summary:
EduTransNet leverages a series of activation functions and carefully designed layers to effectively predict transparency scores. By combining ReLU, Tanh, Leaky ReLU, and Sigmoid activations, the model captures complex patterns in the data, yielding robust, accurate predictions.
Fig. 5
EduTransNet Architecture.
For regression tasks, a linear activation function may be used, while for classification tasks, a softmax or sigmoid activation function is typically employed.
This architecture allows the neural network to learn complex relationships between input features and academic performance metrics, enabling accurate predictions and valuable insights into student learning outcomes.
Architecture of EduTransNet
The architecture of the EduTransNet is designed to capture the intricate relationships between input features and predicted academic performance. It consists of multiple hidden layers, each containing a variable number of neurons. The model architecture is represented as follows:
$$Input\left(Features\right)->HiddenLayer1->HiddenLayer2-n->Output\left(PredictedPerformance\right)$$
Each hidden layer employs a specific activation function to introduce non-linearity and facilitate feature transformation. The number of neurons in each hidden layer and the choice of activation function are determined through experimentation and performance evaluation, ensuring optimal model performance. The equations governing the computation within the proposed neural network model are as follows:
$${h}^{\left(l\right)}=\sigma\left({W}^{\left(l\right)}{h}^{\left(l-1\right)}+{b}^{\left(l\right)}\right)$$
(16)
$$\widehat{y}=\text{s}\text{o}\text{f}\text{t}\text{m}\text{a}\text{x}\left({W}^{\left(L\right)}{h}^{\left(L-1\right)}+{b}^{\left(L\right)}\right)$$
(17)
where:
\({h}^{l}\) is the output of layer l.
\({W}^{l}\) is the weight matrix of layer l.
\({b}^{l}\) is the bias vector of layer l.
σ is the activation function applied element-wise to the weighted sum, introducing non-linearity and capturing complex patterns. SoftMax is the SoftMax function, used in the output layer for multi-class classification tasks, ensuring probabilistic interpretative.
yˆ is the predicted output, representing the model’s estimation of academic performance.
In this research, Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared were selected as the primary evaluation metrics due to their ability to provide a comprehensive assessment of the model’s predictive accuracy and performance. MSE was selected because it penalises larger errors more than smaller ones by squaring the differences between the predicted and actual values. This is particularly important in educational settings where large deviations in predictions (e.g., student performance) could have significant consequences. MSE provides a precise measure of how closely the model’s predictions match the actual outcomes, with lower values indicating more accurate predictions. MAE, on the other hand, measures the average magnitude of errors in predictions, without considering their direction (i.e., whether they are positive or negative). MAE provides a more interpretable metric for understanding the average difference between the predicted and true values, making it useful when analysing the model’s general performance across a wide range of predictions. In educational applications, where even small errors can be meaningful, MAE provides a clearer perspective on overall predictive performance. Finally, R-squared was chosen because it represents the proportion of variance in the dependent variable explained by the independent variables. In other words, it provides insight into how well the model captures the data’s underlying patterns. An R-squared value close to 1, as seen with the EduTransNet model, indicates that the model explains a high proportion of the variability in the target variable, which is a strong indicator of its effectiveness. Compared with benchmarks and other models such as Support Vector Regression (SVR), Linear Regression (LR), and Random Forest Regression (RFR), EduTransNet consistently achieved lower MSE and MAE while maintaining higher R-squared values. This suggests that EduTransNet not only produces more accurate predictions but also captures the complex relationships in the data more effectively than the benchmark models. This combination of metrics ensures a balanced and thorough evaluation of the model’s performance, making it ideal for assessing both accuracy and explanatory power in this study.
Fairness and transparency implementation
Fairness was operationalised through three mechanisms: (1) demographic parity constraints ensuring prediction distributions remain consistent across protected attributes (gender, ethnicity); (2) equalised odds verification ensuring similar true positive and false positive rates across groups; and (3) disparate impact analysis with a threshold of 0.8 (80% rule). Fairness metrics were evaluated using the AI Fairness 360 toolkit, with results indicating demographic parity ratios of 0.92 for gender and 0.89 for ethnicity.
Novelty of EduTransNet
The novelty of EduTransNet extends beyond conventional deep neural network architectures in several key aspects:
(1)
Hybrid Activation Strategy: Unlike standard MLPs that use uniform activation functions, EduTransNet employs a purposefully designed sequence of activation functions (ReLU → Tanh → ReLU → Leaky ReLU → Sigmoid) to capture diverse feature representations at different abstraction levels;
(2)
Embedded Ethical Constraints: The architecture incorporates fairness-aware loss functions that simultaneously optimise for predictive accuracy and demographic parity, a feature absent in traditional regression models;
(3)
Domain-Specific Design: The network topology was engineered explicitly for educational transparency prediction, with layer dimensions calibrated to the inherent complexity of educational data patterns.
(4)
Interpretability Layers: The final sigmoid activation produces bounded outputs (0-100 scale) that directly correspond to interpretable transparency scores, facilitating stakeholder understanding and trust.
Training configuration
Training Configuration
The EduTransNet model was trained using the following hyperparameters.
Optimiser: Adam with default momentum parameters (β₁ = 0.9, β₂ = 0.999).
Learning Rate: 0.001 with exponential decay (decay rate = 0.95 per 10 epochs).
Batch Size: 32.
Epochs: 200 maximum with early stopping (patience = 20 epochs based on validation loss).
Regularization: Dropout rate = 0.3 after each hidden layer; L2 weight decay = 0.0001.