Deep learning framework based on student interest prediction in educational data using HFCTGT-Net

Abstract

Despite significant advancements in educational data mining and Deep Learning (DL), existing student interest prediction models suffer from several critical technical limitations. However, the architecture is intended to process moderately large and complex educational data sets in a simulated environment. Moreover, conventional machine learning and shallow DL approaches often exhibit suboptimal generalization, inefficient feature selection, and inability to integrate multi-source cognitive and behavioral data effectively. To overcome these challenges, this research proposes a novel big data-driven DL framework, termed Hybrid Fusion-Centric Temporal Graph Transformer Network (HFCTGT-Net), designed to deliver robust, scalable, and high-precision student interest prediction. The proposed methodology is structured into a sequence of advanced algorithmic stages within a distributed big data analytics architecture. Initially, Adaptive Cognitive Data Harmonization (ACDH) employs a deep autoencoder-based imputation and normalization strategy to address noise, data sparsity, and missing values in large-scale simulated educational datasets. Subsequently, Hierarchical Deep Feature Abstraction (HDFA) utilizes stacked Variational Autoencoders integrated with attention mechanisms to extract multi-level latent representations, effectively overcoming limitations of shallow feature extraction and preserving complex nonlinear relationships. To address the challenge of modeling dynamic learning behaviors, Sequential Cognitive Behavior Modeling (SCBM) integrates Bidirectional Long Short-Term Memory networks with Transformer architectures, enabling simultaneous capture of long-term temporal dependencies and contextual interactions. For resolving inefficiencies in feature selection and reducing redundant high-dimensional features, a hybrid metaheuristic optimization approach combining Elephant Herding Optimization (EHO) and Social Spider Optimization (SSO) is employed, enhancing feature relevance and computational efficiency. The final prediction stage is realized using the proposed HFCTGT-Net, which incorporates temporal graph learning and fusion-centric attention mechanisms to model complex interdependencies among learners, activities, and contextual factors. Extensive experimental evaluation demonstrates that the proposed model shows better performance than current models, convergence stability, scalability, and computational efficiency, as validated through key performance metrics including precision, recall, F1-score, and processing latency. The experimental results suggest that the proposed model outperforms previous methods in terms of prediction performance. However, the results are based on a synthetic dataset, and future work should evaluate the model on real-world data.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗