From Text to Recommendations: LLM-Guided Interest Tokenization via T5 and Cross-Attention in Generative Models

ETEGRec is a SOTA end-to-end generative recommender system that integrates item tokenization and recommendation generation. However, its latent item tokenization process fails to take into account the high-level semantic information of user interests, which can lead to suboptimal recommendations. To address this, we introduce a brand new interest fusion mechanism that leverages user interest text embeddings to enhance the recommendation process. We propose a recommendation framework that leverages a large language model (LLM) to extract structured user interests from purchase history and item metadata. A multi-layer cross-attention mechanism fuses interest embeddings with item sequences, enhancing semantic relevance modeling. Training uses cached batch processing for consistency, while inference employs asynchronous updates for real-time performance. This approach improves recommendation accuracy with efficient user interest integration.