<![CDATA[This study investigates whether course recommendation and persuasive course promotion can be unified in a fully reproducible pipeline for learning-oriented course promotion. The problem is timely because large language models (LLMs) have rapidly entered education, recommendation, and personalized content generation, yet most educational recommendation studies still optimize ranking accuracy alone, while many promotion-oriented LLM studies emphasize text generation without a rigorous educational recommendation benchmark. We therefore propose an offline end-to-end framework named LGPRec that couples a hybrid recommender with a deterministic template-based explanation planner with an LLM-structured slot design. The recommender integrates collaborative, latent-factor, content-based, and completion-sensitive signals, while the explanation planner uses the same learner profile to produce short persuasive course promotions that remain faithful to the recommended course. Full experimental evaluations were conducted on two public datasets: an IBM course recommendation benchmark and the XuetangX MOOC sequence benchmark. On the IBM dataset, the final model achieved Recall@10, NDCG@10, and MRR@10 of 0.7264, 0.4750, and 0.3962. On XuetangX, the proposed HoTrans sequence model reached 0.4848, 0.3333, and 0.2869. Explanation metrics were evaluated on the IBM benchmark, which provides course text and topic labels; the personalized explanation layer preserved item faithfulness at 1.0000 while raising explicit history mention from 0.0012 to 1.0000 and increasing the unique ratio from 0.0027 to 0.1912. These results show that learning-oriented course promotion can be framed as a recommendation-plus-explanation problem and evaluated rigorously under reproducible offline conditions using public data.]]>
Copyrights © 2027
