Domain adaptive transfer learning for student dropout prediction across online course platforms with limited labelled data
Student dropout prediction in online learning environments suffers from a fundamental tension: platforms with the greatest need for early-warning systems possess the fewest labelled records. This study proposes a three-stage domain-adaptive transfer learning framework built around EduTransformer, a Pre-Layer-Normalisation Transformer encoder that models student activity sequences over a unified cross-platform vocabulary. After training on the Open University Learning Analytics Data Set (OULAD) using a joint masked language modeling and classification task, the proposed framework was fine-tuned for domain-adaptive transfer learning on the unlabeled KDD Cup 2015 dataset across seven strategies, including progressive fine-tuning, SHOT, CORAL, MMD, and CDAN, at four target label scarcity ratios ranging from 1% to 20%. Domain-adaptive pretraining is decisive under extreme label scarcity: at 1% supervision it moves the mean transfer effect from − 0.0102 (without it) to + 0.0025 AUC, while its marginal contribution falls within seed noise from 5% upward. Among transfer strategies, the source-free SHOT method is the most reliable under scarcity: it gives the largest 1% gain ( δ AUC = + 0.0034) and is the only method to survive Benjamini–Hochberg correction at 1%, remaining significant at 1%, 10%, and 20%. Gains become systematic once a modest budget is available, with all six transfer methods significant at 20%.
Authors
- Khalid Fardousse (ORCID: https://orcid.org/0000-0003-3374-5810)
- Zakaria Alj (ORCID: https://orcid.org/0000-0003-0712-4754)
- Anas Bouayad (ORCID: https://orcid.org/0000-0002-3372-1168)
- Youssef Talambouti
Institutions
- Mohammed V University (MA)
- Sidi Mohamed Ben Abdellah University (MA)
Publication Details
- Journal
- Discover Artificial Intelligence
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1007/s44163-026-02746-3
- Primary Topic
- Online Learning and Analytics
- Type
- article
- Field-Weighted Citation Impact
- 0.00