Data-Driven Student Interest Identification for Intelligent Teaching Management Systems in Engineering Colleges

Summary

The Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) protocol integrates quantum-inspired cognitive modeling, data preprocessing, feature selection, and hybrid neural learning to identify engineering students' interests. This protocol presents a reproducible hybrid quantum-inspired machine learning workflow for educational analytics, demonstrated through engineering student interest identification using multidimensional educational data.

Abstract

This protocol presents a quantum-inspired computational framework for identifying engineering students' interests using multidimensional educational and behavioral data. The objective is to provide a systematic and reproducible workflow that integrates data preprocessing, behavioral pattern analysis, feature selection, and hybrid neural learning to support educational decision-making and personalized learning. The workflow begins with Box-Cox Robust IQR Scaling (BCRIS) to preprocess and normalize educational data by reducing semantic noise and improving data quality. Student learning behaviors are subsequently analyzed using the Quantum Cognitive State Model (QCSM), which represents behavioral transitions and evolving learning preferences under uncertain conditions. Relevant features are then selected using the Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm to reduce feature redundancy and identify informative behavioral characteristics. Finally, the Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) integrates quantum-inspired learning with deep neural networks to classify student interests. The protocol was validated using a publicly available educational dataset. Model performance was evaluated using conventional metrics, including accuracy, F1-score, and the area under the receiver operating characteristic curve (AUC), together with Interest Detection Precision (IDP), Quantum Pattern Stability Index (QPSI), and Cross-Domain Adaptability Gain (CDAG). The framework achieved an overall classification accuracy of 96.17% on the evaluated dataset. This protocol provides a reproducible computational workflow for analyzing multidimensional educational data to support intelligent educational management and personalized learning applications.

Introduction

Educational management systems are increasingly used across the education sector, driven by rapid advances in information technology and computer networks. The continued growth of higher education has been accompanied by increased student enrollment, particularly among engineering students, resulting in a substantial increase in educational management data. This growth supports sustainable educational development, accelerates educational informatization, and facilitates the establishment of standardized educational management systems that improve the consistency and efficiency of teaching administration. Machine Learning (ML) provides computational techniques that enable systems to analyze data, learn from patterns, and continuously improve predictive performance without explicit programming^1,2.

Educational quality is a fundamental component of higher education. In recent years, examinations have been widely used to assess engineering students' understanding of course content and their ability to acquire new knowledge, thereby promoting continuous improvements in educational quality. Learner-centered environments can be established by leveraging diverse learning resources available through online educational platforms and applying multiple learning theories to address the needs of different learners. Effective teaching plays a key role in improving learning outcomes among engineering students in universities and colleges^3,4. Furthermore, effective teaching promotes a positive learning environment, supports students' understanding of course content, and enhances academic performance. Educational quality is evaluated using multiple indicators throughout the learning process, including curriculum structure, classroom environment, and teaching content.

Higher education promotes students' intellectual, personal, and social development. At the same time, student engagement—including behavioral, emotional, and cognitive dimensions—and academic persistence have become important determinants of learning outcomes. High levels of student engagement contribute significantly to long-term academic success^5,6. Advances in educational technology have created new opportunities for data-driven educational analytics by modeling complex nonlinear relationships among students' cognitive, behavioral, and academic characteristics using Machine Learning (ML) and Deep Learning (DL) techniques.

A major challenge remains the effective observation and assessment of classroom dynamics while accounting for evolving student behaviors. Traditional approaches for evaluating educational quality and analyzing engineering students' behavior are often time-consuming and susceptible to subjective human judgment. Student inattention during classroom activities adversely affects learning outcomes and academic performance^7,8. However, maintaining sustained attention among engineering students remains challenging because of internal and external distractions and fluctuations in motivation.

Assessing and improving student engagement has become increasingly important with the widespread adoption of online learning, smart classrooms, and other interactive educational technologies. Ensuring inclusion across diverse groups of engineering students further highlights the need for comprehensive frameworks that integrate technology-driven approaches with educational practice while maintaining privacy and ethical standards. In addition, limited model transparency and challenges associated with technology integration^9,10, including methodological bias and educational equity, continue to hinder the development of predictive systems capable of supporting learners through continuous monitoring.

Machine learning approaches have been widely used to address educational challenges such as predicting student risk and dropout rates. These approaches have demonstrated the ability to identify high-risk students, thereby supporting interventions that improve academic achievement¹¹. Artificial Intelligence (AI)-based educational management systems have also been developed to enhance both administrative and academic processes within higher education institutions. Educational datasets commonly include academic records such as course grades, attendance records, and enrollment history¹².

Previous studies have analyzed student behavior from three perspectives: evaluation data, methods, and outcomes^13,14. Prediction models based on information literacy have been developed to analyze the behavioral characteristics of college students. These studies primarily focus on learner behavior assessment using Deep Learning (DL) methods. However, research on intelligent teacher behavior assessment remains at an early stage, limiting progress in this area. Fast Region-based Convolutional Neural Networks (Fast R-CNNs) have been applied to educational datasets to improve classroom activity detection, demonstrating significant performance improvements over traditional methods¹⁵.

Bi-Directional Long Short-Term Memory (Bi-LSTM) models incorporating self-monitoring mechanisms have also improved the accuracy of grade prediction for college students. Similarly, prediction models developed using online learning behavior datasets have demonstrated improved predictive performance^16,17,18. In addition, k-means clustering (KMC) has been used to identify the characteristics of different student groups through cluster analysis, providing insights into student behavior that support improvements in teaching, management, and learning. Deep learning-based strategies have also been proposed to improve the detection and evaluation of classroom teaching behaviors. In these approaches, classroom behavioral features are used to establish relationships between feature representations and hyperplanes using linearly separable starting values for the Support Vector Machine (SVM)^19,20. However, challenges such as occlusion, fine-grained behavioral details, and environmental interference continue to limit the accuracy of engineering student behavior recognition in classroom environments.

Deep learning has also been applied to outcome-based evaluation of university courses, including educational assessment in political and ideological education^21,22. The effectiveness of these prediction models has been demonstrated using publicly available educational datasets. Feed-forward Spiking Neural Networks (FSNNs) have been employed to predict students' academic performance using data collected from online learning platforms and learning management systems. As illustrated in Figure 1, the proposed computational workflow incorporates an educational assessment dataset to identify engineering students' interests through sequential stages of data preprocessing, behavioral pattern analysis, feature selection, and classification. Table 1^{23,24,25,26,27,28,29,30,31,32} summarizes representative methods reported in previous studies, together with the techniques employed and their associated limitations for identifying student interest within learning management systems.

Figure 1: Student interest identification framework. Overall workflow of the proposed engineering student interest identification framework, including data acquisition, preprocessing, behavioral pattern analysis, feature selection, and classification. Please click here to view a larger version of this figure.

Table 1: Student behavior analysis in teaching management. Summary of representative methods, techniques, datasets, evaluation metrics, and reported limitations for student behavior analysis in teaching management. Please click here to download this Table.

The Swin Transformer and Lightweight Feature Pyramid Network (FPN) have been applied to student classroom behavior detection by efficiently processing multi-scale feature maps³³. In addition, the Fruit-Fly-Optimized Adjustable Recurrent Neural Network (PFFO-ARNN) has been used to analyze students' online learning behavior and predict course achievement³⁴. Studies based on student behavior data collected from university learning management systems have also highlighted the inefficiency and time-consuming nature of monitoring classroom behavior with conventional deep learning methods³⁵. Table 2^{36,37,38,39,40,41,42,43,44,45,46,47} summarizes representative classification methods, datasets, evaluation metrics, and reported limitations for intelligent teaching management systems in engineering education.

Table 2: Deep learning-based intelligent teaching management systems for engineering education. Comparison of representative deep learning-based teaching management systems, datasets, evaluation metrics, and reported limitations. Please click here to download this Table.

A previous study⁴⁸ developed a blended teaching quality assessment model based on Deep Belief Networks (DBNs) and proposed a mixed evaluation strategy for assessing teaching quality in universities. The proposed approach outperformed the conventional Genetic Algorithm (GA) in both classification accuracy and execution time, achieving over 95% accuracy in under 5.4 ms. The model was trained using labeled educational datasets and evaluated using an independent testing strategy^49,50. The reported results demonstrated strong predictive performance, with a Mean Absolute Error (MAE) of 0.593% and a Root Mean Square Error (RMSE) of 0.785%.

Recent attention-enhanced architectures, such as KANFormer⁵¹, integrate multi-head self-attention with Kolmogorov-Arnold networks to improve student performance prediction by capturing complex feature interactions. Transformer-based knowledge tracing models have also demonstrated improved prediction accuracy and interpretability by leveraging attention mechanisms to model students' learning trajectories⁵². Although transformer- and attention-based architectures have improved performance on educational prediction tasks by modeling contextual dependencies and long-range feature interactions, these approaches primarily focus on student performance prediction or knowledge tracing rather than uncertainty-aware student interest identification through quantum-inspired cognitive modeling, optimized feature selection, and hybrid meta-learning.

Despite recent advances in machine learning and deep learning for educational analytics, existing methods for identifying engineering students' interests have several limitations. Many conventional approaches focus primarily on prediction accuracy while overlooking the multifaceted nature of student behavior, including evolving learning preferences, behavioral uncertainty, contextual relationships, and semantic interactions. Furthermore, these methods often rely on conventional feature engineering and deterministic learning strategies, which may not adequately capture nonlinear behavioral patterns or adapt to dynamic learning environments. Consequently, a more comprehensive framework integrating robust data preprocessing, intelligent behavioral modeling, adaptive feature selection, and hybrid learning techniques is needed to improve the analysis of complex educational data.

To address these limitations, the proposed framework integrates robust data preprocessing, quantum-inspired behavioral modeling, optimized feature selection, and hybrid quantum-classical neural learning into a unified computational workflow for identifying engineering students' interests. By combining these complementary components, the proposed protocol provides a systematic and reproducible approach for analyzing complex educational data to support personalized learning and educational decision-making.

The objective of this study is to develop and validate the CQEI framework using multidimensional educational and behavioral data for engineering student interest identification. The framework combines robust data preprocessing, quantum-inspired cognitive state modeling, optimized feature selection, and the Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) to analyze students' learning behaviors and facilitate accurate interest identification. Performance is evaluated using standard metrics, including accuracy, F1-score, and the area under the receiver operating characteristic curve (AUC), together with Interest Detection Precision (IDP), Quantum Pattern Stability Index (QPSI), and Cross-Domain Adaptability Gain (CDAG). By integrating these complementary components into a single computational workflow, the proposed framework provides a systematic approach for educational analytics and intelligent educational decision-making.

Protocol

This study used a publicly available, anonymized dataset obtained from the Kaggle repository. No human participants were directly recruited or involved, and no personally identifiable information was collected. Therefore, institutional ethical approval and informed consent were not required. The research tools used in this study are listed in the Table of Materials.

1. Overview of the CQEI framework
The proposed Quantum-Inspired Conceptual Education Intelligence (CQEI) framework integrated data preprocessing, behavioral pattern analysis, feature selection, and classification to identify engineering students' interests using the Teaching Quality Evaluation Dataset obtained from the Kaggle repository. The dataset was first preprocessed using the Box–Cox Robust IQR Scaling (BCRIS) method to reduce semantic noise and normalize the input features. Student behavioral patterns were subsequently analyzed using the Quantum Cognitive State Model (QCSM). The Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm was then applied to identify informative features while reducing feature redundancy by modeling quantum-optimal selection, feedback significance, and transient behavioral variations. Finally, the Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) integrated quantum-inspired learning with deep residual semantic layers to identify students' interests.

The overall workflow of the proposed ML-HQNC framework is illustrated in Figure 2. The protocol comprised six sequential stages, beginning with acquisition of the educational dataset from the Kaggle repository, followed by data preprocessing using Box–Cox Robust IQR Scaling (BCRIS), feature transformation and behavioral representation using the Quantum Cognitive State Model (QCSM), and feature selection using the Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm. The selected features were subsequently used for model training and validation with the Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC), and the framework was finally evaluated using accuracy, F1-score, the area under the receiver operating characteristic curve (AUC), Interest Detection Precision (IDP), Quantum Pattern Stability Index (QPSI), and Cross-Domain Adaptability Gain (CDAG).

Figure 2: Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) framework. Overall architecture of the proposed ML-HQNC framework integrating Box–Cox Robust IQR Scaling (BCRIS) for data preprocessing, the Quantum Cognitive State Model (QCSM) for behavioral pattern analysis, the Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm for feature selection, and hybrid quantum-classical neural learning for student interest classification. Please click here to view a larger version of this figure.

2. Teaching quality evaluation dataset
The proposed framework was developed using the publicly available Teaching Quality Evaluation Dataset obtained from the Kaggle repository. The dataset contained 1,014 educational records with 14 multidimensional features representing academic performance, assessment scores, attendance, instructional characteristics, and student behavioral indicators collected through the Comprehensive Education Management Platform (CEMP). These attributes provided the multidimensional educational information needed to model nonlinear relationships associated with identifying engineering students' interests.

As shown in Figure 3, the dataset was divided into training (80%) and testing (20%) subsets. A total of 800 records were used to train the proposed model, with 200 reserved for performance evaluation. The dataset is publicly available at: https://www.kaggle.com/datasets/ziya07/teaching-quality-evaluation-dataset. The training and testing datasets were subsequently processed using the proposed framework to construct the learning model and improve its predictive capability.

Although the underlying dataset is titled the Teaching Quality Evaluation Dataset, this study aimed to identify engineering students' interests. The educational, behavioral, and instructional attributes contained in the dataset served as input variables for training and evaluating the proposed framework. Consequently, all reported prediction results and performance metrics describe the framework's effectiveness in identifying engineering students' interests rather than evaluating teaching quality. Before data preprocessing, the imported dataset was verified to ensure that all 1,014 records were successfully loaded, the dataset schema matched the expected format, and any missing or duplicate records had been identified and addressed.

3. Data pre-processing using box–cox robust IQR scaling (BCRIS)
The Box–Cox Robust IQR Scaling (BCRIS) method was used to preprocess and normalize the educational dataset by combining the Box–Cox transformation with interquartile range (IQR)-based robust scaling to reduce the influence of extreme values. Symmetry correction together with median-centered normalization generated statistically stable and scalable feature representations suitable for subsequent analysis. Within the proposed educational intelligence framework, BCRIS transformed multidimensional educational, instructional, and behavioral indicators into variance-controlled feature representations. Features including attendance_rate, interactive_sessions_percentage, and related educational variables were normalized to improve the consistency of the training and testing datasets. Detailed mathematical explanations are provided in Supplementary File 1 (Section S1). The BCRIS method provided a quantum-compatible preprocessing strategy that normalized heterogeneous educational data prior to subsequent statistical analysis while preserving the underlying behavioral characteristics required for downstream modeling. Before behavioral analysis, the preprocessing pipeline was verified to ensure that the normalized dataset contained no invalid or missing values, semantic noise had been reduced, and the processed features were suitable for subsequent behavioral modeling.

4. Feature transformation and behavioral representation using the quantum cognitive state model (QCSM)
The Quantum Cognitive State Model (QCSM) was implemented to model the dynamics of engineering students' interests using quantum-inspired probabilistic representations. Student interest was represented as an overlap of multiple conceptual states that reflected the hierarchy of cognitive concepts, uncertainty, conceptual overlap, simultaneous engagement with multiple topics, hesitant or indecisive learning behaviors, and the probability of shifts in learning preferences. The multidimensional indicators from the Engineering Teaching Quality Assessment dataset, including Student_Avg_Score, Attendance_rate, and Student_Feedback_Rating, were represented as amplitude-weighted cognitive transitions within a normalized quantum state space. The QCSM combined quantum-state transitions with interference-based conceptual alignment to generate quantum-inspired cognitive representations that reflected the stability and evolution of student interest over time. Detailed mathematical formulations are provided in Supplementary File 1 (Section S2). The QCSM modeled probabilistic cognitive states to represent knowledge acquisition, conceptual change, learning engagement, and subject interest through normalized quantum-state representations. Behavioral representations were verified across all student records to ensure that the extracted cognitive-state features accurately captured learning behaviors and topic transitions.

5. Feature selection using Quantum-Enhanced Swarm–Whale Optimization Feature Selector (QES-WOFS) algorithm
The Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm was used to identify the optimal feature subset and determine high-information behavioral characteristics from multidimensional educational data. Quantum-inspired search mechanisms enhanced the integration of global and local search strategies for feature optimization. Educational variables, including interactive_sessions_percentage, assignments_time_percentage, and course_completion_rate, were evaluated according to their correlation with the target outcome. Candidate feature subsets were modeled using quantum probability representations, while a quantum tunneling mechanism enabled probabilistic transitions beyond local optima. Feature importance was estimated using marginal likelihood to quantify each feature's contribution to classification performance and adaptive swarm states. The QES-WOFS algorithm selected stable and behaviorally meaningful features, improving robustness to noise, semester-to-semester variation, and periodic behavioral fluctuations.

To overcome the limitations associated with using either Whale Optimization Algorithm (WOA) or Particle Swarm Optimization (PSO) independently, the QES-WOFS algorithm combined PSO-based global exploration with WOA-based local exploitation. A quantum tunneling mechanism further enabled probabilistic transitions beyond local optima, thereby increasing population diversity, reducing feature redundancy, and improving convergence toward an optimal feature subset for identifying engineering student interest.

Detailed mathematical formulations are provided in Supplementary File 1 (Section S3). The QES-WOFS algorithm combined quantum probabilistic feature encoding, a PSO-adapted velocity strategy, and a whale-inspired spiral search mechanism to identify the most informative educational features. Hyperparameters were empirically optimized using the training dataset. Candidate values were evaluated through systematic experimentation, and the combination producing the highest validation performance was selected. Specifically, the Box–Cox transformation parameter (λ = 0.25), the contrastive loss margin (m = 1.0), and the quantum tunneling coefficients (α = 0.7 and β = 0.3) were selected based on validation accuracy, F1-score, and convergence behavior. Before classifier training, the feature selection results were verified to ensure that the QES-WOFS algorithm had selected the optimal feature subset and removed redundant or low-importance features.

6. Model training and validation using the meta-learning hybrid quantum-classical neural classifier (ML-HQNC)
The Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) was implemented by combining deep neural networks with quantum-inspired computations to identify engineering students' interests. Multi-head quantum attention mechanisms and deep residual semantic layers captured complex nonlinear relationships within multidimensional educational data represented using variational quantum circuits (VQCs). Behaviorally derived features, including student_average_grade, attendance_rate, and student_feedback_rating, were projected into a quantum state space, where normalized inputs were transformed into amplitude-based quantum representations to model overlapping and probabilistic engagement patterns. Normalized student behavioral feature vectors were calculated according to the equations provided in Supplementary File 1 (Section S4). The ML-HQNC subsequently classified and identified engineering students' interests.

The ML-HQNC classifier employed an episodic meta-learning strategy during training. One randomly selected meta-learning task from the training dataset was assigned to each episode. While maintaining the class distribution, the training data were divided into a support set (80%) and a query set (20%). The support set was used for task-specific parameter adaptation through the inner optimization loop, whereas the query set evaluated the adapted model and calculated the meta-loss used to update the shared parameters. A total of 100 meta-learning tasks were generated during each training cycle. Five gradient update steps with an inner learning rate of 0.001 were used for inner-loop adaptation, whereas the Adam optimizer with an outer learning rate of 0.0005 optimized the meta-parameters. Random task sampling was repeated for every training epoch to expose the model to diverse behavioral patterns and improve generalization to previously unseen student profiles. The loss estimation for quantum contrastive alignment is described in Supplementary File 1 (Equations 41–45). Model training was verified to ensure successful convergence without computational errors and that the training and validation loss curves had stabilized over successive epochs.

Results

The proposed ML-HQNC framework was evaluated against Fast R-CNN, Bi-LSTM, CNN, and XGBoost using conventional classification metrics, including Accuracy, Precision, Recall, F1-score, and the area under the receiver operating characteristic curve (AUC), together with the proposed evaluation metrics Interest Detection Precision (IDP), Quantum Pattern Stability Index (QPSI), and Cross-Domain Adaptability Gain (CDAG). The mathematical definitions of these metrics are provided in Supplementary File 1 (Section S5).

To assess the robustness of the proposed framework, all experiments were repeated five times using different random initializations. The reported performance values are presented as the mean ± standard deviation across the five runs. Statistical significance between the proposed framework and the baseline models was evaluated using a paired t-test, with p < 0.05 considered statistically significant.

As summarized in Supplementary Table 1, simulation experiments were conducted using the Teaching Quality Evaluation Dataset containing 1,014 records. A total of 800 records were used for model training, whereas 200 records were reserved for testing. The proposed ML-HQNC framework and comparison models were implemented in the Jupyter Notebook environment using Python. The Teaching_Quality attribute served as the original target variable in the publicly available Teaching Quality Evaluation Dataset and consisted of two classes, Good and Excellent. These labels were used as the prediction targets for training and evaluating the ML-HQNC framework. The resulting prediction categories are summarized in Table 3. Supplementary Table 2 summarizes the validation results of the Box–Cox Robust IQR Scaling (BCRIS) preprocessing stage. The preprocessing pipeline removed missing and duplicate records, reduced the influence of outliers, normalized feature distributions, and improved semantic and behavioral consistency before feature transformation and model training.

Table 3: Target variable of the Teaching Quality Evaluation Dataset. Description of the original Teaching_Quality target variable and its classification categories (Good and Excellent) used for model training and evaluation. Please click here to download this Table.

Supplementary Table 3 presents representative features generated by the Quantum Cognitive State Model (QCSM). The extracted features describe student learning behavior, including cognitive state, engagement, uncertainty, topic transitions, and behavioral consistency, and served as the input to the subsequent QES-WOFS feature selection stage. Supplementary Table S4 lists the final features selected by the QES-WOFS algorithm together with their feature categories, relative influence levels, and functional roles within the proposed framework. The characteristics of the extracted multidimensional behavioral features are summarized in Supplementary Table 5. Supplementary Table 6 presents representative edge structures generated during behavioral representation for the two prediction classes (Good and Excellent).

As shown in Figure 4 and Table 4, the proposed ML-HQNC framework achieved the highest F1-score (87.9%) among the evaluated methods. The corresponding F1-scores for Fast R-CNN, Bi-LSTM, CNN, and XGBoost were 77.9%, 80.2%, 82.8%, and 85.8%, respectively. Statistical significance was assessed using a two-sided independent t-test (p < 0.05).

Figure 5 and Table 5 summarize the Interest Detection Precision (IDP) results. The proposed ML-HQNC framework achieved an IDP of 89.3%, compared with 79.11%, 82.12%, 84.6%, and 87.10% for Fast R-CNN, Bi-LSTM, CNN, and XGBoost, respectively. Figure 6 and Table 6 present the Quantum Pattern Stability Index (QPSI) results. The proposed framework achieved a QPSI value of 92.7%, whereas Fast R-CNN, Bi-LSTM, CNN, and XGBoost achieved 81.1%, 84.11%, 87.4%, and 90.9%, respectively. Figure 7 and Table 7 summarize the Cross-Domain Adaptability Gain (CDAG) results. The proposed ML-HQNC framework achieved the highest CDAG value (0.95), compared with Fast R-CNN (0.78), Bi-LSTM (0.83), CNN (0.80), and XGBoost (0.86). As shown in Figure 8 and Table 8, the proposed ML-HQNC framework achieved the highest classification accuracy (96.17%). The corresponding accuracies for Fast R-CNN, Bi-LSTM, CNN, and XGBoost were 85.10%, 87.21%, 90.14%, and 93.19%, respectively. Figure 9 and Table 9 present the computational time required by each method. The proposed ML-HQNC framework required 9.02 ms, whereas Fast R-CNN, Bi-LSTM, CNN, and XGBoost required 31.2 ms, 27.5 ms, 22.4 ms, and 17.6 ms, respectively. Figure 10 presents the receiver operating characteristic (ROC) analysis. The proposed ML-HQNC framework achieved an AUC of 0.944, indicating a high true-positive rate and a low false-positive rate across the evaluated classification threshold.

Figure 4: F1-score comparison across different dataset sizes. Comparison of the F1-score achieved by ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 4: F1-score comparison across different dataset sizes. Comparison of the F1-score achieved by ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

Figure 5: Interest Detection Precision (IDP) comparison across different dataset sizes. Comparison of Interest Detection Precision (IDP) achieved by ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 5: Interest Detection Precision (IDP) comparison across different dataset sizes. Comparison of Interest Detection Precision (IDP) achieved by ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

Figure 6: Quantum Pattern Stability Index (QPSI) comparison across different dataset sizes. Comparison of the Quantum Pattern Stability Index (QPSI) achieved by ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 6: Quantum Pattern Stability Index (QPSI) comparison across different dataset sizes. Comparison of the Quantum Pattern Stability Index (QPSI) achieved by ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

Figure 7: Cross-Domain Adaptability Gain (CDAG) comparison across different dataset sizes. Comparison of the Cross-Domain Adaptability Gain (CDAG) achieved by ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 7: Cross-Domain Adaptability Gain (CDAG) comparison across different dataset sizes. Comparison of the Cross-Domain Adaptability Gain (CDAG) achieved by ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

Figure 8: Classification accuracy comparison across different dataset sizes. Comparison of the classification accuracy achieved by ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 8: Classification accuracy comparison across different dataset sizes. Comparison of the classification accuracy achieved by ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

Figure 9: Computational time comparison across different dataset sizes. Comparison of the computational execution time of ML-HQNC and competing models across different dataset sizes. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Error bars represent one standard deviation. Lower values indicate shorter computational execution time. Statistical significance was evaluated using a two-sided independent t-test (p < 0.05). Please click here to view a larger version of this figure.

Table 9: Computational time comparison across different dataset sizes. Comparison of the computational execution time of ML-HQNC and competing models. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Lower values indicate shorter computational execution time. Please click here to download this Table.

Figure 10: Area under the receiver operating characteristic curve (AUC) comparison. Comparison of the AUC achieved by ML-HQNC and competing models for engineering student interest identification. The reported AUC values correspond to a single experimental run; therefore, error bars and statistical significance analyses are not presented. Please click here to view a larger version of this figure.

Supplementary Figure 1 presents the correlation matrix of representative behavioral variables. Strong positive correlations were observed between Avg_Score and Course_Completion_Rate (0.81) and between Attendance_Rate and Course_Completion_Rate (0.75). Moderate correlations were observed between Avg_Score and Course_Completion (0.72) and between Attendance_Rate and Time_Assignment_Percentage (0.74). Supplementary Figures 2 and 3, together with Table 10, summarize the training and testing accuracy maps. The proposed ML-HQNC framework achieved a training accuracy of 96.1% and a testing accuracy of 93.8%, exceeding the performance of the comparison methods. Supplementary Figures 4 and 5, together with Table 11, summarize the convergence analysis. The proposed ML-HQNC framework achieved the lowest training loss (0.21) and testing loss (0.26), with stable convergence throughout the optimization process.

Table 12 summarizes the overall performance of the proposed ML-HQNC framework and the comparison methods across all evaluation metrics. The proposed framework achieved the highest performance for Accuracy, F1-score, AUC, IDP, QPSI, and CDAG while requiring the shortest computational time among the evaluated methods. Table 13 presents the component-wise ablation analysis of the proposed framework. Progressive incorporation of the BCRIS preprocessing stage, the QCSM behavioral representation, the QES-WOFS feature selection algorithm, and the complete ML-HQNC classifier yielded incremental improvements in classification performance, demonstrating each component's contribution to the overall framework.

Table 10: Training and testing accuracy comparison. Comparison of the training and testing accuracy achieved by ML-HQNC and competing models. The reported values correspond to a single experimental run. Please click here to download this Table.

Table 11: Training and testing loss comparison. Comparison of the training and testing loss achieved by ML-HQNC and competing models. The reported values correspond to a single experimental run. Please click here to download this Table.

Table 12: Overall performance comparison. Comparison of ML-HQNC and competing models using multiple evaluation metrics. Results are reported as mean ± standard deviation over five independent experimental runs (n = 5). Please click here to download this Table.

DATA AVAILABILITY:
The original Teaching Quality Evaluation Dataset used in this study is publicly available from the Kaggle repository (https://www.kaggle.com/datasets/ziya07/teaching-quality-evaluation-dataset). Access to the dataset may require user authentication and is subject to Kaggle's terms of use. To support reproducibility while complying with the licensing conditions of the original dataset, the study-generated resources—including the synthetic dataset, source code, implementation and configuration files, feature descriptions, preprocessing workflow, and supporting documentation required to reproduce the reported experiments—have been deposited in an openly accessible Zenodo repository: https://doi.org/10.5281/zenodo.21393413.

Figure 3: Teaching Quality Evaluation dataset feature collection. Structure and feature composition of the Teaching Quality Evaluation Dataset used for engineering student interest identification. Please click here to view a larger version of this figure.

Supplementary Figure 1: Feature correlation matrix. Correlation matrix showing pairwise relationships among representative educational and behavioral variables used by the proposed ML-HQNC framework. Please click here to download this file.

Supplementary Figure 2: Training accuracy comparison. Comparison of training accuracy achieved by ML-HQNC and competing models over the training process. Please click here to download this file.

Supplementary Figure 3: Testing accuracy comparison. Comparison of testing accuracy achieved by ML-HQNC and competing models over the testing process. Please click here to download this file.

Supplementary Figure 4: Training loss analysis. Comparison of training loss achieved by ML-HQNC and competing models during model optimization. Please click here to download this file.

Supplementary Figure 5: Testing loss analysis. Comparison of testing loss achieved by ML-HQNC and competing models during model evaluation.  Please click here to download this file.

Supplementary Table 1: Simulation parameters and experimental settings. Summary of the simulation parameters, implementation environment, dataset partitioning, and experimental settings used to evaluate the proposed ML-HQNC framework. Please click here to download this file.

Supplementary Table 2: Validation results of the Box–Cox Robust IQR Scaling (BCRIS) preprocessing stage. Comparison of representative dataset characteristics before and after preprocessing, including missing values, duplicate records, outlier handling, feature normalization, and data consistency. Please click here to download this file.

Supplementary Table 3: Representative features generated by the Quantum Cognitive State Model (QCSM). Representative behavioral features generated by the QCSM following data preprocessing, including cognitive state, engagement, uncertainty, topic transition, and behavioral consistency features. Please click here to download this file.

Supplementary Table 4: Features selected by the Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm. Final feature subset selected by the QES-WOFS algorithm, together with the corresponding feature categories, relative influence levels, and functional roles within the proposed framework. Please click here to download this file.

Supplementary Table 5: Teaching quality prediction categories. Description of the prediction categories and representative behavioral characteristics used by the proposed ML-HQNC framework. Please click here to download this file.

Supplementary Table 6: Margin sensitivity analysis. Summary of the margin sensitivity analysis used during model evaluation. Please click here to download this file.

Supplementary File 1: Supplementary Methods, Mathematical Formulations, Evaluation Metrics, and Supplementary Tables. This file contains detailed mathematical formulations for the Box–Cox Robust IQR Scaling (BCRIS), Quantum Cognitive State Model (QCSM), Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS), and Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) described in the main protocol. It also includes definitions of the evaluation metrics, simulation parameters, experimental settings, and Supplementary Tables S1–S6 supporting the results presented in the main manuscript. Please click here to download this file.

Discussion

The proposed ML-HQNC framework demonstrated improved performance in identifying engineering student interest by integrating quantum-inspired behavioral modeling, adaptive feature selection, and meta-learning into a unified computational workflow. Compared with Fast R-CNN, Bi-LSTM, CNN, and XGBoost, the framework consistently achieved higher performance across conventional classification metrics and the proposed behavioral evaluation metrics. These findings suggest that combining quantum-inspired cognitive state representation with adaptive feature optimization improved the framework's ability to model complex educational behaviors while maintaining robust classification performance. The low computational time and the small difference between training and testing performance further indicate stable optimization and good generalization.

The proposed evaluation metrics complemented conventional classification measures by providing additional information on behavioral modeling. Interest Detection Precision (IDP) quantified the accuracy of interest identification, Quantum Pattern Stability Index (QPSI) evaluated the stability of the quantum-inspired behavioral representations, and Cross-Domain Adaptability Gain (CDAG) assessed the ability of the framework to generalize across different data partitions. Together, these measures provided a broader assessment of model performance than conventional prediction metrics alone.

The component-wise ablation analysis (Table 13) further demonstrated the contribution of each stage of the proposed workflow. Incorporating the Box–Cox Robust IQR Scaling (BCRIS) preprocessing stage improved data consistency by reducing the influence of outliers and generating normalized feature representations. The Quantum Cognitive State Model (QCSM) transformed multidimensional educational variables into behavioral representations that captured uncertainty, engagement, and evolving learning preferences. Subsequent feature selection using the Quantum-Enhanced Swarm–Whale Optimization Feature Selection (QES-WOFS) algorithm reduced feature redundancy while retaining informative behavioral characteristics. Finally, integrating these components within the Meta-Learning Hybrid Quantum-Classical Neural Classifier (ML-HQNC) improved classification performance and generalization compared with the baseline models.

The proposed protocol is intended for engineering student interest identification using multidimensional educational and behavioral data collected from digital learning environments. The computational workflow can support personalized learning, academic advising, curriculum planning, and intelligent educational management by providing a systematic approach for analyzing student learning behaviors and identifying evolving interests. The framework is particularly suitable for structured educational datasets containing behavioral, academic, and contextual variables.

Successful implementation depends on the availability of high-quality educational data and adequate computational resources. Before analysis, educational datasets should be examined for missing values, duplicate records, and inconsistencies because data quality directly influences preprocessing and downstream behavioral modeling. Stable feature representations also depend on appropriate preprocessing, feature transformation, and hyperparameter optimization. During model training, convergence should be monitored through the training and validation loss curves, and optimization parameters should be adjusted when necessary to maintain stable learning behavior.

Although the framework demonstrated effective performance using the evaluated dataset, several limitations should be considered. The protocol was validated using a single publicly available educational dataset, and additional evaluation using larger, multi-institutional, and more diverse educational datasets is needed to further assess generalizability. In addition, the hybrid quantum-inspired learning framework requires greater computational resources than conventional machine learning approaches because of its preprocessing, optimization, and hybrid learning components. The framework also depends on well-annotated educational data, and performance may be affected by incomplete, noisy, or highly imbalanced datasets.

Future work will focus on extending the framework to finer-grained multi-class student interest identification, incorporating additional educational data modalities such as textual feedback, temporal learning behaviors, and real-time learning management system activity logs, and evaluating the protocol across diverse engineering education environments. Further optimization of the computational workflow may also improve scalability and facilitate deployment in large-scale educational management systems.

In summary, the proposed ML-HQNC framework provides a reproducible computational workflow for engineering student interest identification using multidimensional educational data. By integrating robust preprocessing, quantum-inspired behavioral modeling, adaptive feature selection, and hybrid quantum-classical neural learning, the framework achieved improved classification performance while maintaining stable optimization and efficient computation. These characteristics support its potential application in intelligent educational analytics and engineering education management.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Baseline Models (Fast R-CNN, Bi-LSTM, CNN, XGBoost)	Open-source implementations	Fast R-CNN (2015); TensorFlow/Keras implementation v2.16; XGBoost v2.1.1	Benchmark models used for comparative performance evaluation.
BCRIS Preprocessing Module	Developed in this study	Version 1.0	Custom preprocessing module implementing Box-Cox transformation and robust IQR normalization.
Jupyter Notebook	Project Jupyter	Version 7.2.2	Interactive development environment used for implementation, training, and testing.
ML-HQNC Classifier	Developed in this study	Version 1.0	Proposed hybrid quantum-classical neural network classifier with meta-learning framework.
Python	Python Software Foundation	Version 3.11.9	Programming language used for model development and experimentation.
PyTorch	PyTorch Foundation	Version 2.5.1	Deep learning framework used for model implementation and training.
QCSM Model	Developed in this study	Version 1.0	Proposed Quantum Cognitive State Model for behavioral pattern representation.
QES-WOFS Algorithm	Developed in this study	Version 1.0	Proposed Quantum-Enhanced Swarm–Whale Optimization Feature Selection algorithm.
Scikit-learn	Scikit-learn Developers	Version 1.6.1	Machine learning library used for preprocessing, feature selection, and evaluation.