A lightweight multi-scale deep separable network for early dropout prediction in MOOCs: balancing accuracy and educational deployability

Massive Open Online Courses (MOOCs) have expanded access to higher education, yet high dropout rates persist. Early identification of at-risk learners is essential for timely support, but current predictive models often prioritize technical accuracy over practical applicability in educational contexts. This study investigates a lightweight multi-scale deep learning model for early dropout prediction in MOOCs, designed for use on typical school and university hardware. The proposed Multi-scale Depthwise Separable Convolutional Neural Network (Multi-scale DSCNN) features decision-tree-based data cleaning to reduce noise, parallel multi-scale feature learning to capture relationships in learner behavior, and a symmetric architecture that reduces computational requirements. Evaluated on the KDD Cup 2015 dataset, the model achieved 91.5% accuracy, 97.7% recall, and 95.1% F1-score. High recall ensures few at-risk students are missed, while the efficient design allows near-real-time prediction on common devices. Balancing accuracy with deployability (0.16M parameters, 1.8M FLOPs, CPU-only inference, 85 MB RAM), this approach enables practical integration of dropout prediction in educational settings, supporting timely and equitable interventions, (97.7% recall, 2.3% miss rate), and is suitable for standard school computers (runs on Intel i5-8250U with 1.82 ms per sample inference time).