Optimized Deep Learning Techniques to Predict Learners’ Online Performance

Existing models for predicting students’ academic performance show limitations which result in their inability to make accurate predictions. The system achieved operational efficiency together with precise predictions, which created problems for its application in extensive online educational platforms. The system fails to predict accurately because it depends solely on basic dataset attributes, which prevents it from understanding complex user behavior patterns. The research introduces a refined Deep Neural Network (DNN) classifier to predict student performance. The research connects these study gaps with results obtained from UK Open University Learning Analytics Dataset (OULAD) which includes Social Science and STEM courses. The study introduced new features that track student behaviors through different activities which include Engagement, Total Pre-Course Activities, Average, Studying, Discussing, Examining, Working, and Total Post-Course Activities. The research team selected their most predictive features through Particle Swarm Optimization (PSO), which resulted in a feature reduction from 40 to 12. The research team optimized the hyperparameters through Grid Search, which tested learning rates between 0.0001 and 0.1 and assessed three hidden unit configurations: 64, 128, and 256. The proposed model achieved accuracies of 90% (Social Science) and 88% (STEM), which showed better performance than previous studies that used the same datasets with lower computational requirements. The results enable educators to perform personalized educational interventions which will improve student results during online courses