LIGER - A hybrid retrieval model launched by Meta AI and other institutions
LIGER is a hybrid retrieval model developed by Meta AI and other institutions, combining the advantages of generative retrieval and dense retrieval. LIGER uses a generative retrieval module to generate a finite set of candidate items, and then uses dense retrieval to rank and...
What is LIGER?
LIGER, a hybrid retrieval model developed by Meta AI and other institutions, combines the advantages of generative and intensive retrieval. LIGER uses a generative retrieval module to generate a finite set of candidate items, and then uses intensive retrieval to rank and optimize these items. This retains the storage and inference efficiency advantages of generative retrieval, improving recommendation performance, particularly when handling cold-start items. LIGER predicts the semantic ID and text representation of the next item by taking its semantic ID and item text representation as input, effectively bridging the performance gap between generative and intensive retrieval and providing an efficient and accurate solution for recommendation systems.
LIGER's main functions
- Efficiently generate candidate projectsBased on the generative retrieval module, a limited set of candidate items can be quickly generated, reducing the number of items that need to be searched intensively and improving retrieval efficiency.
- Optimize candidate project rankingDense retrieval is used to sort and optimize the generated candidate items, ensuring the accuracy and relevance of the recommendation results and improving the user experience.
- Handling cold start projectsIt effectively generates and recommends cold-start projects, solves the recommendation problem caused by the lack of historical interaction data for new projects, and enhances the robustness of the recommendation system.
- Combining semantic informationBased on semantic IDs and project text representations, we can delve deeper into the semantic information of projects, enhance our understanding of project content, and improve the accuracy of recommendations.
LIGER's technical principles
- Generative retrieval moduleBased on the Transformer generative model, a sequence of semantic IDs for candidate items is generated according to the user's interaction history and item characteristics. Using semantic IDs to represent the semantic information of items reduces the need for individual embedding of each item.
- Dense search moduleDense retrieval techniques are applied to the generated candidate item set. The similarity between candidate items and target items is calculated using the textual representation and semantic ID of the items. The candidate items are then ranked and optimized, and the most relevant items are selected for recommendation.
- Multi-objective optimizationDuring training, the objective functions of generative retrieval and dense retrieval are optimized simultaneously. By minimizing the prediction error of generative retrieval and maximizing the similarity of dense retrieval, the model performs well in both generating candidate items and optimizing ranking.
- Integrating semantic ID and text representationUsing semantic IDs and item text representations as input features, and based on the collaborative work of the encoder and decoder, the embedded representations of items are generated, while the semantic IDs and text representations of the next item are predicted, thus achieving an effective fusion of generative and dense retrieval.
LIGER's project address
- arXiv technical paper:https://arxiv.org/pdf/2411.18814
Application scenarios of LIGER
- e-commerce platformBased on users' shopping history and preferences, generate and optimize a set of product candidates to provide personalized recommendations, thereby improving the shopping experience and conversion rate.
- Content PlatformBased on users' reading habits, generate and sort a candidate set of relevant articles, recommend the most interesting articles, and increase users' reading time and platform stickiness.
- social mediaAnalyze users' social relationships and interactions, generate and sort a potential friend candidate set, recommend people who may become friends, enhance the social experience and expand the network.
- Online education platformBased on students' learning history and interests, generate and optimize a set of course candidates, recommend the most suitable courses, and improve learning outcomes and teaching quality.