What Is LIGER
LIGER is a hybrid retrieval model introduced by Meta AI and other institutions, combining the advantages of generative retrieval and dense retrieval. LIGER uses a generative retrieval module to generate a limited set of candidate items, and ranks and optimizes the candidate items based on dense retrieval, retaining the advantages of generative retrieval in storage and inference efficiency, and improving recommendation performance, especially in handling cold-start items. LIGER is based on taking semantic IDs and item text representations as input, predicting the semantic ID and text representation of the next item, effectively bridging the performance gap between generative and dense retrieval, and providing an efficient and accurate solution for recommendation systems.
LIGER's Main Features
- Efficient Candidate Item Generation: Based on the generative retrieval module, it quickly generates a limited set of candidate items, reducing the number of items that require dense retrieval and improving retrieval efficiency.
- Optimized Candidate Item Ranking: It uses dense retrieval to rank and optimize the generated candidate items, ensuring the accuracy and relevance of recommendation results and improving user experience.
- Handling Cold-Start Items: It effectively generates and recommends cold-start items, solving the recommendation challenge caused by new items lacking historical interaction data and enhancing the robustness of the recommendation system.
- Combining Semantic Information: Based on semantic IDs and item text representations, it deeply mines the semantic information of items, enhances understanding of item content, and improves recommendation accuracy.
LIGER's Technical Principles
- Generative Retrieval Module: A Transformer-based generative model that generates semantic ID sequences of candidate items based on users' interaction history and item features. It represents item semantic information based on semantic IDs, reducing the need for separate embeddings for each item.
- Dense Retrieval Module: It applies dense retrieval techniques to the generated candidate item set, using item text representations and semantic IDs to calculate the similarity between candidate items and target items, ranks and optimizes candidate items, and selects the most relevant items for recommendation.
- Multi-Objective Optimization: During training, it simultaneously optimizes the objective functions of generative retrieval and dense retrieval. Based on minimizing the prediction error of generative retrieval and maximizing the similarity of dense retrieval, it enables the model to perform well in both generating candidate items and optimizing ranking.
- Fusing Semantic IDs and Text Representations: It takes semantic IDs and item text representations as input features, and based on the collaborative work of the encoder and decoder, generates item embedding representations while predicting the semantic ID and text representation of the next item, achieving effective fusion of generative and dense retrieval.
LIGER's Project Address
- arXiv Technical Paper：https://arxiv.org/pdf/2411.18814
LIGER's Application Scenarios
- E-commerce Platforms: Based on users' shopping history and preferences, it generates and optimizes product candidate sets, provides personalized recommendations, and improves shopping experience and conversion rates.
- Content Platforms: According to users' reading habits, it generates and ranks relevant article candidate sets, recommends the most interesting articles, and increases users' reading time and platform stickiness.
- Social Media: It analyzes users' social relationships and interactions, generates and ranks potential friend candidate sets, recommends people who may become friends, and enhances social experience and network expansion.
- Online Education Platforms: Based on students' learning history and interests, generate and optimize a candidate set of courses, recommend the most suitable courses, and improve learning outcomes and teaching quality.