Core Concepts
Generative search and recommendation leverage powerful generative language models to directly generate relevant documents or items in response to user queries or profiles, revolutionizing traditional information retrieval methods.
Abstract
This paper provides a comprehensive survey on the emerging paradigm of generative search and recommendation. It first summarizes the previous machine learning-based and deep learning-based paradigms in search and recommendation, which approach the tasks as discriminative matching problems. In contrast, the generative paradigm formulates the tasks as generation problems, aiming to directly generate the target documents or items.
The survey then abstracts a unified framework for generative search and recommendation, consisting of four key steps: query/user formulation, document/item identifiers, training, and inference. Within this framework, the paper categorizes and analyzes the existing works on generative search and recommendation, highlighting their strengths, weaknesses, and unique challenges.
For generative search, the paper discusses various document identifiers, including numeric IDs, titles, n-grams, codebooks, and multiview identifiers. It also examines the training and inference processes, including generative and discriminative training, as well as free generation and constrained generation.
For generative recommendation, the paper focuses on the user formulation, which incorporates task descriptions, user's historical interactions, user profiles, context information, and external knowledge. It also reviews the different item identifiers, such as numeric IDs and textual metadata.
The survey further delves into the comparison between generative search and recommendation, identifies open problems in the generative paradigm, and envisions the next information-seeking paradigm that could emerge from the advancements in large language models.
Stats
"With the information explosion on the Web, search and recommendation are foundational infrastructures to satisfying users' information needs."
"Search can be formulated as a matching between queries and documents, and recommendation can be considered a matching between users and items."
Quotes
"Embracing generative search and recommendation brings new benefits and opportunities for the field of search and recommendation. In particular, 1) LLMs inherently possess formidable capabilities, such as vast knowledge, semantic understanding, interactive skills, and instruction following. These inherent abilities can be transferred or directly applied to search and recommendation, thereby enhancing information retrieval tasks. 2) The tremendous success of LLMs stems from their generative learning. A profound consideration to apply generative learning to search and recommendation, fundamentally revolutionizing the methods of information retrieval rather than only utilization of LLMs. 3) LLMs-based generative AI applications, such as ChatGPT, are gradually becoming a new gateway for users to access Web content. Developing generative search and recommendation could be better integrated into these generative AI applications."