Description

This paper introduces a novel approach to courselet recommendations for the Quantinar peer-to-peer educational platform by leveraging advanced language models and hybrid recommender architectures, highliting the release of a new publicly available dataset. Our method addresses challenges posed by highly correlated datasets through two complementary techniques. First, CourseletGCN employs LLaMA-based embeddings and a Planar Maximally Filtered Graph (PMFG) to construct a semantic graph that effectively captures intricate relationships among courselets. Second, SearchEnsemble models user behavior with a gated recurrent unit (GRU) to harness sequential browsing patterns and search term embeddings. Extensive benchmarking against state-of-the-art models using metrics such as Hit Ratio, Precision, Recall, MAP, NDCG, and Gini Index demonstrates notable improvements in recommendation diversity. A real-world case study on the Quantinar platform further confirms the system’s ability to enhance user engagement and experience, while integrated MLOps practices ensure continuous deployment and optimization. Moreover, the introduced dataset, comprising 414 courselets and over 40,000 validated interactions at the time of writing, offers a valuable resource for future research and development in recommender systems