LLM‐Enhanced Semantic Information Alignment for ID Recommendation

ABSTRACT

Identity‐based Recommendation (IDR) plays a crucial role in modern personalized recommendation systems by leveraging users' identity information and historical interactions to model long‐term user preferences. However, existing IDR methods face three critical challenges: difficulty in fully capturing latent user preference, limited representational capacity, and poor adaptability to cold‐start scenarios. To address these issues, we propose the LLM‐Enhanced Semantic Information Alignment for ID Recommendation (ESIA‐IDR) framework, which systematically integrates semantic features generated by large language models (LLMs) into ID‐based recommendation. ESIA‐IDR first employs LLMs to deeply perceive unstructured textual data, such as user reviews and product descriptions, thereby effectively mining users' latent interest patterns. An independent pretraining mechanism is then applied to optimize the training dynamics of semantic and ID embeddings, mitigating inconsistencies in convergence behavior. Additionally, a feature discretization strategy is adopted to enhance semantic feature compatibility with traditional recommendation pipelines. During the recommendation phase, semantic‐enhanced features and collaborative signals are jointly modeled to achieve robust performance in cold‐start environments. ESIA‐IDR consists of four key components: Semantic and Interest Perception, Semantic Alignment, Discrete Feature Construction, and Recommendation Integration. By combining pretraining, dimensionality reduction, and similarity‐based bucketing, the framework achieves effective alignment between semantic features and ID feature spaces, bridging the semantic‐structural gap and improving user preference modeling accuracy. Extensive experiments on multiple real‐world datasets demonstrate that ESIA‐IDR achieves a 2.21% improvement in NDCG and a 1.48% improvement in Hit Rate compared to the best‐performing baselines, validating its effectiveness in enhancing recommendation performance.