Generative Recommendation: Towards Personalized Multimodal Content Generation

Traditional recommender systems primarily retrieve items from a pre-defined item corpus for personalized recommendations. However, this a retrieval-based paradigm faces two major constraints: 1) the existing item corpus with human-generated items may not adequately satisfy users' diverse information needs, and 2) users often rely on passive and noisy feedback (e.g., clicks) to refine the recommendations. To overcome the limitations, the rapid advancement of AI-Generated Content (AIGC) presents significant potential: 1) generative AI promotes the creation of personalized multimodal content as new items to satisfy users' information needs, and 2) large language models reduce the efforts of users to express information needs proactively in natural language. In this light, we propose a novel Generative Recommender paradigm named GeneRec with two objectives: 1) generating personalized multimodal content, and 2) integrating proactive user instructions to guide content generation. To achieve the objectives, GeneRec introduces an AI generator to personalize multimodal content generation and leverages user instructions to obtain users' information needs. Specifically, GeneRec first employs an instructor to pre-process users' instructions and traditional feedback (e.g., clicks) to extract generation guidance. Following the guidance, we develop the AI generator with an AI editor to edit existing items and an AI creator to create new items, respectively. Lastly, we study the feasibility of implementing the AI creator in the fashion domain, showing promising results. Furthermore, to ensure the trustworthiness of the generated items, we emphasize various trustworthiness checks such as fairness and safety checks.