A BERT-Based Course Recommendation Framework Using Weighted Knowledge Gaps

Misalignment between higher education curricula and labor-market requirements complicates course selection for students. Traditional course recommendation systems rarely target occupational outcomes and typically treat course completion as a binary event, disregarding how well a student performed. We propose a recommendation framework based on Sentence-BERT (SBERT) that represents a student’s competency deficit relative to a target occupation as a directional vector, obtained by subtracting an attainment-weighted student profile from the occupational requirement vector. Grade-Weighted Pooling constructs the profile so that verified proficiency determines each completed course’s contribution; competency deficits then emerge through subtraction rather than through an explicit remediation rule. We compared nine configurations combining three text representations (TF-IDF, Word2Vec, and SBERT) with three profile-construction strategies, using 2310 courses and 530 occupational requirement statements from Thailand, evaluated against a blinded expert-annotated reference standard over 20 student profiles. SBERT with Grade-Weighted Pooling achieved the highest expert-rated relevance among the examined configurations (Precision@10 = 0.880, NDCG@10 = 0.689). Because annotation was performed over a pooled candidate set drawn from the outputs of the nine configurations, these values measure ranking quality within that pool rather than retrieval effectiveness over the full 2310-course catalogue. Inter-annotator agreement on the reference standard was high (Krippendorff’s α = 0.914). Results derive from a single institutional curriculum and an offline evaluation.