An Attentive Deep Learning Framework for the Prediction of Academic Performance of Students in Virtual Learning Environments

Virtual Learning Environments (VLEs) and Massive Open Online Courses (MOOCs) are two trends that have emerged due to the usage of new technology in education. However, without adequate monitoring, the students will suffer from online education’s inability to assess student achievement. In this research, the Attentive Pelican Optimized Dense Gated Recurrent Unit (APDGRU) network is introduced. In order to eliminate the information that is required for prediction, one-hot encoding is used. DenseNet and Gated Recurrent Unit (GRU) are two deep learning models included in the proposed network design, along with another method for carefully evaluating the relative contributions of different elements. The dataset parameters, which mostly consist of categorical data like the student’s personal information, are fed into the suggested model. The primary variables in the dataset that have the largest impacts on student performance are found using the DenseNet model. The proposed model’s GRU unit deals with temporal data, such as how long students engage in VLE-based learning and interaction. The attention mechanism is crucial for helping the model recognize the best traits, which it learns throughout training. The attention mechanism’s integration enhances the model’s interpretability and accuracy by concentrating on the most pertinent characteristics affecting academic success. The proposed model parameters enable the creation of the best predictions considering formal metric maximization and interpretability. By minimizing the logarithmic loss function, the model’s internal parameters are familiar with the Pelican Optimization Algorithm (POA). POA modifies model weights to ensure the suggested model performs at its best by reducing errors, accelerating the convergence and avoiding overfitting. Furthermore, POA aids in the selection of pertinent characteristics and the avoidance of unnecessary data, resulting in a more robust and generalized model. The implementations are carried out in the Python platform, and the evaluations are done using the Open University Learning Analytics (OULAD) dataset. The overall accuracy of 97.18% proves the model’s benefit over other existing models.