Large Language Model-empowered Course Recommendation with Learning Interest-Goal Contrastive Learning

An efficient course recommendation system plays a critical role in improving individual learning efficiency and advancing educational equity. However, existing course recommendation systems often overlook the presence of noise in learning sequences and fail to fully capture the inter-course dependencies. To address these limitations, we propose a novel course recommendation model that integrates large language models (LLMs) and course dependency structures to enhance sequence modeling. Specifically, we first employ prompt-engineered LLMs for learning courses denoising. Subsequently, we construct a course dependency graph from large-scale learning behavior data, and apply graph convolutional networks (GCNs) to learn course semantics. These representations are then fed into a Transformer to capture learners’ dynamic learning interests. Furthermore, we leverage LLMs to infer learners’ long-term goals and introduce a contrastive learning strategy to align these goals with sequential learning interests, thereby further improving recommendation accuracy. Extensive experiments on two real datasets demonstrate that the proposed model outperforms other baselines. The simulated noise experiment also highlights the superior performance of the model against noisy interactions.