A multilingual heterogeneous knowledge graph framework for curriculum–vacancy–MOOC alignment: edge-weight regression with heterogeneous graph transformers for educational recommendation

Abstract

Digitalisation and artificial intelligence have widened the gap between competencies certified by university curricula and skills demanded by employers. MOOCs could close this gap, but existing recommenders rarely connect what a programme teaches, what the market requires, and what online courses develop within a single semantic structure. We built a multilingual heterogeneous knowledge graph (KG) that integrates university programmes, IT vacancies from hh.kz, and Coursera MOOCs in English, Russian, and Kazakh. A property-graph ontology was formalised and instantiated automatically. Skills were aligned with a fine-tuned multilingual Sentence Transformer (paraphrase-multilingual-mpnet-base-v2), and agglomerative clustering formed skill communities. Recommendation was cast as edge-weight regression over a competency-to-course relation and solved with GraphSAGE, HAN, and HGT encoders. On a held-out 20% test graph, HGT achieved the best ranking quality (NDCG@10 = 0.714; with short meta-paths NDCG@10 = 0.987, RMSE = 0.108), outperforming GraphSAGE and HAN. Applied to the “Information Systems” programme of a Kazakhstani university, the system expanded the market-aligned competency profile by +36.2% with five top-ranked courses and by +170% with the full catalogue. Edge-weight regression on a cross-domain heterogeneous KG produces front-loaded, actionable course recommendations for curriculum alignment. All datasets, models, graph snapshots, and code are released for reproducibility.