STUDY AND ANALYSIS OF STUDENTS ACADEMIC PERFORMANCE USING ARTIFICIAL INTELLIGENCE

Abstract

This thesis explores multiple innovative machine learning and deep learning approaches to accurately predict student performance, providing educational institutions with valuable insights for improving academic fineness and identifying students at high risk for academic challenges. The research introduces four distinct models: Transient Search Capsule Network dependent on Deep Autoencoder (TSCNDE), Gannet Hunt Long Short Term Memory (Gannet Hunt-LSTM), Discriminability-Enhanced Transformer Architecture, and Explainable Deep Learning based Knowledge Distillation Framework (EDL_KDF). Each approach aims to enhance prediction precision, recall, accuracy, and other performance metrics, contributing to the understanding of how students' academic achievement can be forecasted more effectively. Initially, the TSCNDE model uses a novel approach combining a deep autoencoder and transient search capsule network to anticipate the execution of student. The model demonstrated high accuracy and precision by processing data from Open University Learning Analytics Dataset (OULAD), with to the execution metrics achieving up to 99.2% accuracy and 99.8% precision. By focusing on crucial features and the influence of students' online activity, the method provides a reliable tool for assessing student success.Furthermore, the Gannet Hunt-LSTM model integrates a unique hybrid optimization approach by combining the behavior of the Gannet and Harris Hawk for optimal feature selection and classifier learning. This method minimizes overfitting issues and reduces information loss, resulting in exceptional prediction metrics such as 99.9% F-measure, 99.5738% recall, and 99.99% precision. The model leverages deep learning and LSTM networks to understand effect of students' social networking services utilization on academic performance. Additionally, the Discriminability-Enhanced Transformer Architecture aims to improve feature-based prediction by using a hybrid DL approach to analyze students' academic performance. The model utilizes distance- and density-based outlier detection and multi-scale entropy techniques to enhance the prediction process. By extracting features and eliminating data inconsistencies, the approach excels in comparison to existing models and supports educational institutions in better understanding student learning behaviors. Finally, the EDL_KDF model leverages a deep explainable distillation approach with SHapley Additive Explanations (SHAP) to estimate endangered students as well as provide tactics for early intervention. By combining knowledge distillation framework with improved data augmentation techniques, the model achieves 99.6% accuracy and 99.21% precision over the OULAD dataset. The work supports educational institutions' efforts to improve teaching quality, enhance students' academic success, and optimize Virtual Learning Environments (VLEs).