Conference Agenda

While curricular flexibility in engineering programs enhances student autonomy, it introduces significant complexity in academic planning for students and institutions. Ineffective decision-making regarding course sequences can hinder academic progression and increase the risk of dropout. Consequently, analyzing curricular trajectories is essential for supporting student success. This study explores the application of deep learning models to model course enrollment sequences at the School of Engineering of the [ANONIMIZED INSTITUTION], using a dataset of 11,729 students between 2013 and 2025. We compare two neural network architectures—Bidirectional Long Short-Term Memory (bi-LSTM) and Bidirectional Encoder Representations from Transformers (BERT)—against a K-Nearest Neighbors (KNN) baseline in a masked course prediction task. Our results demonstrate that BERT significantly outperforms the alternative models. This superiority is attributed to BERT's ability to model concurrent course enrollments within a single academic period through positional encoding. These findings suggest that BERT-based architectures are highly effective for understanding complex curricular trajectories, offering a robust foundation for course recommendation systems and student progression analytics. This approach opens new avenues for identifying "bottlenecks," predicting timely graduation, and mitigating dropout risk in flexible educational environments.