CASPER-O: An Occupation-Aware Knowledge-Guided Framework for Career-Aligned Course Recommendation

Occupation-aligned course recommendation requires connecting occupational requirements with relevant educational content. CASPER-O addresses this task through occupation-specific retrieval, ESCO-based queries (CASPER Query), and expert-grounded encoder adaptation (CASPER-CL). The corpus contained 11,938 occupation–course associations and 4536 unique courses across ten IT occupations; the expert-annotated benchmark comprised 1200 pairs across eight occupations. Ordinal Krippendorff’s alpha was 0.685. The data partitions were globally course-disjoint. Two complementary protocols evaluated fully judged held-out rankings and retrieval of known held-out relevant courses from larger occupation-specific candidate pools without treating unjudged courses as nonrelevant. Relative to pretrained encoders using CASPER Query, adaptation increased Held-out Relevant Recall@50 from 0.2417 to 0.3208 for MPNet and from 0.3417 to 0.4292 for BGE, while nDCG@10 changed by +0.0037 and −0.0153, respectively. With query–course scores fixed, candidate restriction reduced candidate pairs by 75.53% while retaining all 46 known held-out relevant pairs. Sensitivity analysis found nof target mapping uniformly superior across backbones and metrics. Exact paired sign-flip tests did not attain significance after Holm correction; non-significance does not establish equivalence. Overall, the results distinguish candidate-space effects from model- and metric-dependent effects of query construction and expert-grounded adaptation.