Leveraging Time-Series Foundational Model for Student Dropout Prediction: A Comparative Study

Online courses, ranging from fully online degrees to Massive Open Online Courses (MOOCs), have expanded access to higher education but continue to suffer from extreme dropout rates. Accurate and timely prediction of disengagement is crucial for enabling targeted interventions, yet most existing student dropout prediction (SDP) systems rely on domain-specific models trained on individual course datasets, with limited generalisability across platforms and content types. This paper investigates whether UniTS, a domainagnostic Transformer for time-series data, can be effectively transferred to the task of dropout prediction in sparse, imbalanced educational logs. UniTS was originally designed to handle diverse time-series problems through a unified architecture and dynamic attention mechanisms, without taskspecific engineering. We evaluate multiple adaptation strategies, from supervised training from scratch, to full fine-tuning, to prompt-tuning, and inference-only. Three public datasets of markedly different scale, semantic richness, and class imbalance are considered: XuetangX, KDDCup15, and Unitelma. Results show that a five-epoch fine-tuning of the supervised UniTS checkpoint achieves a mean AUC of 0.944 and F1 of 0.753, outperforming both classical machine learning baselines (e.g., Logistic Regression, Random Forest) and deep sequence models (e.g., GRU-AE) across all datasets. Prompt-tuning retains much of the predictive power while reducing computational cost, highlighting the model's flexibility for resource-constrained deployments. These findings demonstrate that temporal representations learned in unrelated domains can transfer effectively to educational settings, offering a reusable, high-performance foundation model for dropout prediction.