Meta AI researchers introduced a new AI model called LIGER, which integrates the advantages of dense retrieval and generative retrieval, improves the performance of generative recommender systems, and solves the problems of traditional recommender systems in terms of computational resources, storage, and processing of cold-start items, providing new ideas for building recommender systems. On the background of the project, recommender system is important for contacting users and related content, etc. Conventional dense retrieval method requires a large amount of computational resources and storage, and the scalability is limited, and the emerging generative retrieval method can reduce the storage requirements but has performance problems, especially in the cold-start project performance is obvious. In the project introduction, Meta AI jointly launched the LIGER model with multiple organizations, mixing the advantages of the two retrieval methods, using generative retrieval to generate candidate sets and other item representations, and then borrowing dense retrieval techniques to refine them, using bi-directional Transformer encoders and generative decoders to achieve a balance between efficiency and accuracy, reducing computational requirements and generalizing to unseen items. In terms of performance, LIGER consistently outperforms existing state-of-the-art models such as TIGER, UniSRec, etc. in benchmark dataset evaluations such as Amazon Beauty, Sports, Toys, and Steam, e.g., it outperforms the cold-start item Recall@10 score comparisons on Amazon Beauty, Steam datasets, and as the number of the number of retrieval candidates of the generative method increases, its performance gap with dense retrieval narrows, reflecting adaptability and efficiency.
Paper address: https://arxiv.org/abs/2411.18814