Exploring Dual-Attention Neural Networks for Academic Risk Detection: A Literature Review

Abstract

Student performance prediction is a critical task in educational data mining, particularly in digital learning environments where timely intervention can improve learning outcomes. Traditional models often struggle to process the dynamic nature of student behavior across various digital platforms. Recent advancements in deep learning, especially the integration of attention mechanisms within artificial neural networks, have shown significant promise in capturing complex temporal and behavioral patterns from student activity data. This review paper explores the application of attention-based artificial neural networks for predicting student performance, with a focus on a dual-attention model that evaluates both time-based and feature-based importance. The reviewed studies leverage data from platforms such as Moodle and BookRoll, incorporating features like attendance, reading behavior, and assignment submission to identify at-risk students. In addition to summarizing the main study that inspired this project, the paper critically analyzes several other research contributions involving recurrent neural networks, hybrid CNN-LSTM models, knowledge distillation, and fairness-aware learning systems.The review highlights the strengths, limitations, and interpretability of these models, while also identifying key research trends and open challenges. The paper concludes with recommendations for future research aimed at building transparent, adaptable, and ethically responsible learning analytics systems.