Association rules and classification for recommending among equivalent courses on Indian MOOC platforms: evidence from SWAYAM and NPTEL
Abstract Learners on India's national MOOC platforms often find one subject offered by several instructors, institutions, and course lengths, with no evidence on how earlier runs of each offering turned out. We assembled an open, run-level dataset from public pages of NPTEL and the SWAYAM course explorer: 18,344 SWAYAM course offerings from ten national coordinators and 10,347 NPTEL course runs (2016–2026) with published enrolment, exam registration, and certification counts. Outcomes were defined relative to each semester cohort, and every predictor was limited to information available before a run opens. Apriori mining produced 232 non-redundant rules linking course attributes and run history to high or low certification yield; 90.3% of the 226 rules that could be tested kept a confidence of at least 0.60 on 2025–2026 runs that were not used for mining. Eight classifiers were compared under grouped cross-validation and a temporal hold-out. A hybrid model that passes rule matches to gradient-boosted trees reached 84.4% accuracy (AUC 0.928) on 2,955 hold-out runs, against 50.1% for a majority baseline, and a rule-list classifier reached 83.2%. When used to rank equivalent courses taught by different instructors, the hybrid model ordered 75.5% of 110 hold-out pairs correctly and put the best course first in 73.6% of 72 choice sets; ranking by popularity did no better than chance. Run history carried most of the signal, and accuracy fell to 70.5% for courses never seen in training. The dataset and code are released for reuse.
Authors
- Iram Naim (ORCID: https://orcid.org/0000-0002-2255-9839)
- Mohd Moin Kamran
Institutions
- M.J.P. Rohilkhand University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11488802/v1
- Primary Topic
- Online Learning and Analytics
- Type
- preprint