Abstract
With the rapid growth of digital educational platforms, early prediction of student academic performance has become increasingly important for timely risk identification and effective instructional support. This study investigates the computational efficiency and predictive performance of various deep learning recurrent architectures for student performance classification based on time-series data of learning activities. Experiments are conducted using the Open University Learning Analytics Dataset (OULAD), where student interactions with the learning management system are transformed into multidimensional time series combining dynamic behavioral features and static demographic attributes. Multiple recurrent, hybrid, and attention-based models are evaluated in an early prediction setting across different stages of course progression using standard classification metrics and statistical significance testing. In addition to predictive performance, the computational cost of each architecture is systematically assessed in terms of training time, model complexity, memory usage, and storage requirements. The results indicate that while predictive performance across architectures is largely comparable, their computational demands differ substantially, highlighting the importance of cost–performance trade-offs. Based on these findings, the study recommends favoring simpler recurrent models that achieve competitive accuracy with significantly lower computational overhead for deployment in real-world educational systems.