A Survey of Generative Search and Recommendation in the Era of Large Language Models (2404.16924v1)

Abstract: With the information explosion on the Web, search and recommendation are foundational infrastructures to satisfying users' information needs. As the two sides of the same coin, both revolve around the same core research problem, matching queries with documents or users with items. In the recent few decades, search and recommendation have experienced synchronous technological paradigm shifts, including machine learning-based and deep learning-based paradigms. Recently, the superintelligent generative LLMs have sparked a new paradigm in search and recommendation, i.e., generative search (retrieval) and recommendation, which aims to address the matching problem in a generative manner. In this paper, we provide a comprehensive survey of the emerging paradigm in information systems and summarize the developments in generative search and recommendation from a unified perspective. Rather than simply categorizing existing works, we abstract a unified framework for the generative paradigm and break down the existing works into different stages within this framework to highlight the strengths and weaknesses. And then, we distinguish generative search and recommendation with their unique challenges, identify open problems and future directions, and envision the next information-seeking paradigm.

A Comprehensive Survey of the Generative Paradigm in Search and Recommendation Systems

Overview of Generative Search and Recommendation

This paper presents a detailed survey of generative search and recommendation, focusing on their emergence as a significant paradigm in the information retrieval and recommender systems domain. The authors provide an extensive assessment of how generative models, particularly LLMs, are being utilized to innovate traditional search and recommendation processes. These processes traditionally relied on matching queries with documents or users with items using discriminative models. The transformation to a generative approach introduces a fundamentally different method by focusing on generating relevant documents or items directly in response to user queries or profiles.

Unified Framework for Generative Search and Recommendation

The paper introduces a unified framework that encapsulates the essence of generative search and recommendation. This framework comprises four critical stages:

1. Input Formulation: Transforming user queries or profiles into a text form that can be processed by generative models.
2. Identifier Generation: Leveraging models to generate identifiers that can uniquely represent documents or items.
3. Training: Utilizes existing query-document or user-item pairs to train the generative models, ensuring they can effectively generate correct identifiers from given inputs.
4. Inference: Deploying the trained model to generate identifiers for new queries or users, which are then mapped back to documents or items.

This framework simplifies the understanding of how generative models are applied in these systems and highlights the transition from traditional methods that focus on ranking and retrieval to ones that emphasize content generation.

Challenges and Unique Aspects

The survey places a strong emphasis on the novel challenges and unique aspects introduced by the generative paradigm:

• Identifier Systems: The choice of identifiers (numeric IDs, document titles, coded representations) significantly impacts the functionality and effectiveness of the generative approach. For instance, while numeric IDs may offer distinctiveness, they lack semantic meaning which complicates their generation by LLMs.
• Training Issues: Generative models require substantial training to effectively produce accurate and relevant results. This involves not only mapping inputs to correct outputs but also ensuring that the model can generalize well across unseen data.
• Inference Mechanisms: The complexity of generating content that matches user queries or profiles during inference includes ensuring the fidelity and relevance of the generated outputs, often necessitating constrained generation techniques to avoid producing invalid results.

Comparative Discussion

A significant contribution of this paper is the comparative analysis between generative search and generative recommendation systems. Although they share a common framework, distinct differences arise from their core objectives and the nature of the data they handle. For instance, generative search deals with shorter and more direct queries, while recommendation systems often handle complex user profiles involving historical data, which complicates the input formulation process.

Future Directions and Open Problems

The paper does not shy away from addressing the nascent nature of the generative paradigm and its associated challenges. It discusses the potential for these systems to evolve into fully generative formats, where systems might not just retrieve but actually create new content dynamically in response to queries. Issues such as updating model knowledge, handling multimodal inputs, and enhancing the in-context learning capabilities of these models are presented as critical areas for future research.

Conclusion

In sum, this survey serves as a critical document for researchers and practitioners in the field, providing both a comprehensive overview of the current state of generative search and recommendation and insightful discussions about potential innovations and improvements for future systems. The careful examination of how generative models can transform the information retrieval landscape opens up numerous opportunities for advancing the technology further.