Abstract
This article presents an approach to developing digital educational systems for analyzing and predicting the academic performance of university students using machine learning methods.Learning analytics enables the consideration of individual student characteristics and supports decision-making in educational planning.The study was conducted using a dataset containing academic records of 2,423 students collected over five semesters.Several types of models were evaluated, including linear, tree-based, ensemble, and neural network algorithms.A hybrid neural architecture is proposed that combines embedding-based representations of categorical features, a conditional attention mechanism, and LSTM-based sequential modeling to process students' academic history.The experimental results demonstrate performance differences among the tested models.Baseline approaches, including linear, tree-based, kernel-based, and neural network methods, achieved accuracy in the range of approximately 0.71 to 0.86.The proposed Attention-LSTM model achieved an accuracy of 0.89, with F1-macro and F1-weighted scores of 0.88 and 0.89, respectively.The results indicate that hybrid neural architectures can effectively capture patterns in students' academic data.The proposed approach can be applied in digital educational systems to monitor and analyze student learning dynamics.