This presentation addresses the problem of early prediction of student academic performance for AI-driven educational decision support, with a focus on two tasks: Semester GPA (SGPA) prediction and Graduation Classification prediction. The study investigates a complete AI pipeline, from educational data construction and preprocessing to predictive modeling and evaluation, using heterogeneous data collected from Vietnam, and public benchmark datasets. The proposed framework systematically compares traditional machine learning methods with SOTA deep learning architectures, including sequential, graph-based, and generative models. To address challenges of educational data, such as limited samples, class imbalance, uncertainty, and heterogeneous attributes, some novel models are developed: NeutroDL, NeutroGNT, LATCGAd, and AWG-GC. These models integrate neutrosophic representation learning, graph neural networks, attention mechanisms, and adversarial learning to improve prediction robustness and generalization. The resulting prediction models provide the technical foundation for intelligent early-warning, learning analytics, and AI-assisted decision support systems for students, lecturers, and university administrators.

Keywords: AI for Education; Educational Data Mining (EDM); Learning Analytics (LA); Academic Performance Prediction; Deep Learning; Graph Neural Networks (GNN); Hybrid Learning Models; Early Warning Systems.