A fisher kernel vector learning framework with Bahdanau-Adapted rectified deep neural networks for predicting student academic performance

Educational Data Mining (EDM) utilizes data mining, machine learning (ML), and deep learning (DL) techniques to extract valuable insights from educational environments. Although predicting student performance is crucial for enhancing academic achievements, existing prediction models often suffer from class imbalance issues and fail to select optimal features, leading to high computational complexity and reduced accuracy. To address these challenges, this study proposes a novel framework termed Fisher Kernel Vector Machine Learning and Bahdanau Attention Rectified Deep Learning (FKVML-BARDL). The proposed model operates in two primary phases: (1) preprocessing using a Fisher Kernel Globality–Locality Vector Machine Learning model to balance classes while preserving local and global feature properties via Kirchhoff matrices, and (2) classification using Rectified Deep Neural Networks combined with a Bahdanau Attention Adaptable function to select relevant features. The performance of the FKVML-BARDL model was evaluated using a comprehensive student academic performance dataset. Experimental results indicate that the proposed method outperforms existing baseline models across all evaluation metrics, achieving a 17% improvement in precision and recall, an 11% increase in accuracy, and a 31% reduction in computational time.