Large language model enhanced embeddings for knowledge aware recommender systems

Abstract Recommender systems are central to modern digital platforms, yet interaction-only approaches often struggle with sparsity, cold-start users/items, and limited interpretability in domains where rich textual and semantic signals are available. We present a hybrid framework that uses large language models (LLMs) to construct and enrich a domain-specific knowledge graph (KG) from item-side text, and then injects the resulting semantic structure into downstream recommendation models. The proposed pipeline follows a two-phase design. In Phase 1, we induce a movie-domain KG from MovieLens metadata and textual signals, align it to external resources, and evaluate KG quality through complementary intrinsic strategies covering exact matching, structural diagnostics, and embedding-based semantic similarity. In Phase 2, we learn KG-derived embeddings (including translational and bilinear variants) and integrate them into a broad set of state-of-the-art recommenders spanning collaborative filtering, feature-interaction models, and graph-based models. Across models and metrics, KG-augmented representations yield consistent improvements in top- K ranking quality (e.g., Recall@ K , NDCG@ K , Precision@ K ). Beyond end-to-end accuracy, we analyze how intrinsic KG quality signals relate to downstream recommendation gains, and we distill practical guidance on when LLM-generated KGs are most beneficial. Overall, this provides a reusable pipeline and empirical evidence that coupling LLM-based knowledge induction with structured KG reasoning can improve both recommendation effectiveness and interpretability.