A Deep Learning-Based Framework for At-Risk Student Detection Using Educational Data Mining with a Focus on Learning Health

Abstract

Detection of at-risk students through Educational Data Mining (EDM) with emphasis on learning health seeks to determine students who are likely to underachieve or drop out by monitoring trends in their learning behavior, performance indicators, and attendance levels. Current models in this Area are plagued by several issues, such as poor capability to capture deep relational features, low generalization through shallow structures, inadequate treatment of imbalanced and high-dimensional educational data, and the absence of adaptive optimization for varied learning patterns. To overcome these constraints, this study introduces a new deep hybrid model named Harris Hawk Optimization based Capsule Deep Residual and Dense Network model (HHO-CapDeReD-Net), which integrates four strong architectures Capsule Network, Deep Neural Network, ResNet-50, and DenseNet-121 to extract multi-level features from encoded student data. The model also incorporates Harris Hawk Optimization (HHO) to fine-tune hyperparameters for optimal learning performance dynamically and minimize overfitting. By concentrating on learning health, the model continuously monitors and predicts educational risk so that early intervention and tailored help can be ensured. The intended HHO-CapDeReD-Net model shows higher prediction accuracy and stability in pinpointing students who are at risk, thus furthering more successful educational interventions and student retention.