Only content delivered through the publication boundary on this date is included.
arXiv This work proposes the End-to-End Dual Dynamic (ED²) recommender, the first LLM-based sequential recommender to adopt a dual dynamic index mechanism that unifies user/item index generation and sequential recommendation into a single jointly optimized LLM-backbone pipeline, complemented by a multi-grained token regulator (m-GTR) and instruction-tuning tasks for high-order user-item interaction patterns, achieving average improvements of 19.62% in Hit-Rate and 21.11% in NDCG over static-index SOTA LLM recommenders on three public datasets.
This work proposes the End-to-End Dual Dynamic (ED²) recommender, the first LLM-based sequential recommender to adopt a dual dynamic index mechanism that unifies user/item index generation and sequential recommendation into a single jointly optimized LLM-backbone pipeline, complemented by a multi-grained token regulator (m-GTR) and instruction-tuning tasks for high-order user-item interaction patterns, achieving average improvements of 19.62% in Hit-Rate and 21.11% in NDCG over static-index SOTA LLM recommenders on three public datasets.
This work proposes the End-to-End Dual Dynamic (ED²) recommender, the first LLM-based sequential recommender to adopt a dual dynamic index mechanism that unifies user/item index generation and sequential recommendation into a single jointly optimized LLM-backbone pipeline, complemented by a multi-grained token regulator (m-GTR) and instruction-tuning tasks for high-order user-item interaction patterns, achieving average improvements of 19.62% in Hit-Rate and 21.11% in NDCG over static-index SOTA LLM recommenders on three public datasets.
This work proposes the End-to-End Dual Dynamic (ED²) recommender, the first LLM-based sequential recommender to adopt a dual dynamic index mechanism that unifies user/item index generation and sequential recommendation into a single jointly optimized LLM-backbone pipeline, complemented by a multi-grained token regulator (m-GTR) and instruction-tuning tasks for high-order user-item interaction patterns, achieving average improvements of 19.62% in Hit-Rate and 21.11% in NDCG over static-index SOTA LLM recommenders on three public datasets.