PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative Recommendation: Towards Next-generation Recommender Paradigm

Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, Tat-Seng Chua

arXiv:2304.03516v2cs.IR

TL;DR

Retrieval-based recommenders may not satisfy diverse information needs and mainly rely on passive feedback. The paper proposes GeneRec, which uses an instructor with an AI editor and creator to generate, repurpose, or create personalized content from instructions and feedback. GeneRec provides a framework for retrieval, repurposing, and creation, while micro-video creation remains less satisfactory and motivates further work.

  • Problem

    Retrieval-based recommenders may lack human-generated content that satisfies diverse needs and rely on inefficient passive feedback.

  • Method

    GeneRec uses an instructor to process instructions and feedback, an AI editor to repurpose items, and an AI creator to create new items.

  • Results

    GeneRec supports item retrieval, repurposing, and creation, and the paper explores its feasibility for micro-video generation.

  • Takeaways & Limitations

    Generative recommendation should combine user instructions, feedback, content generation, and fidelity checks across recommendation domains.

  • Takeaways & Limitations

    Personalized micro-video creation is less satisfactory because single-turn instructions omit details and datasets lack sufficient videos, facts, and knowledge.

Abstract

from arXiv · show

Recommender systems typically retrieve items from an item corpus for personalized recommendations. However, such a retrieval-based recommender paradigm faces two limitations: 1) the human-generated items in the corpus might fail to satisfy the users' diverse information needs, and 2) users usually adjust the recommendations via inefficient passive feedback, e.g., clicks. Nowadays, AI-Generated Content (AIGC) has revealed significant success, offering the potential to overcome these limitations: 1) generative AI can produce personalized items to satisfy users' information needs, and 2) the newly emerged large language models significantly reduce the efforts of users to precisely express information needs via natural language instructions. In this light, the boom of AIGC points the way towards the next-generation recommender paradigm with two new objectives: 1) generating personalized content through generative AI, and 2) integrating user instructions to guide content generation. To this end, we propose a novel Generative Recommender paradigm named GeneRec, which adopts an AI generator to personalize content generation and leverages user instructions. Specifically, we pre-process users' instructions and traditional feedback via an instructor to output the generation guidance. Given the guidance, we instantiate the AI generator through an AI editor and an AI creator to repurpose existing items and create new items. Eventually, GeneRec can perform content retrieval, repurposing, and creation to satisfy users' information needs. Besides, to ensure the trustworthiness of the generated items, we emphasize various fidelity checks. Moreover, we provide a roadmap to envision future developments of GeneRec and several domain-specific applications of GeneRec with potential research tasks. Lastly, we study the feasibility of implementing AI editor and AI creator on micro-video generation.

1 INTRODUCTION

Traditional recommenders may lack content that satisfies diverse personalized needs and rely on inefficient passive feedback. GeneRec addresses these limitations by combining user instructions and feedback with generative AI for personalized content generation.

  • Motivation: GeneRec targets two limitations of retrieval-based recommendation: insufficient corpus content for diverse needs and reliance on inefficient passive feedback.The paper motivates generation and instruction-guided interaction as responses to these limitations.
  • Motivation: AIGC enables personalized content generation, while LLMs help users express information needs through natural-language instructions.These capabilities motivate new recommender objectives centered on generation and instruction integration.
  • GeneRec: GeneRec adds an AI-generator loop that uses user instructions and feedback to generate content for direct recommendation or item-corpus ranking.Instructions may be textual or multimodal, including images, videos, audio, and natural language.
  • GeneRec: Its instructor processes instructions and feedback into generation guidance, while an AI editor repurposes existing items and an AI creator makes new items.The modules instantiate personalized editing and creation within the proposed paradigm.
  • Contributions: The paper emphasizes fidelity checks and outlines applications, future research directions, and a micro-video feasibility study.The contributions include trustworthiness perspectives and domain-specific research tasks.

2 GENERATIVE RECOMMENDER PARADIGM

GeneRec complements retrieval by using instructions and feedback to generate personalized content that can be directly recommended or ranked with existing items. The paradigm also emphasizes fidelity, evaluation, and a roadmap spanning interaction, generation, and recommender algorithms.

  • Overview: GeneRec introduces two objectives: automatically repurposing or creating items with generative AI and integrating rich user instructions.These objectives complement the traditional retrieval-based recommender paradigm.
  • Overview: Generated items may be directly exposed to users or added to the item corpus for ranking alongside human-generated items.Direct exposure is considered when users clearly request generated content or repeatedly reject human-generated items.
  • User instructions: Instructions can be textual or multimodal, while feedback such as clicks supplies implicit preferences that instructions may omit.The generator preprocesses both signals to learn personalized information needs.
  • Content generation: Content generation includes repurposing existing items and creating new items, followed by refinement, relevance assessment, and fidelity checks.Fidelity checks target accurate, fair, and safe generated content.
  • Fidelity checks: GeneRec must address bias, privacy, and safety because generated items may perpetuate harmful biases, leak personal information, or harm users.These concerns motivate checks tailored to recommendation scenarios.
  • Roadmap and evaluation: The roadmap covers user-system interaction, content generation, and recommender algorithms, with evaluation including item-side and user-side setups.The paper frames these as future development directions for GeneRec.

3 DEMONSTRATION

GeneRec is instantiated with an instructor, AI editor, and AI creator that convert user signals and knowledge into personalized edited or newly created items. The paper applies the paradigm across domains and identifies instruction-following and implicit-preference modeling as open research challenges.

  • 3.1 Instructor: The instructor initializes generation and transforms multimodal instructions and historical feedback into guidance signals.It may use multi-turn interaction and decide whether generation should begin.
  • 3.2 AI editor: The AI editor combines guidance, an existing corpus item, and Web facts or knowledge to produce an edited item tailored to user preferences.External facts and knowledge support accuracy, safety, legality, and updating with new events or regulations.
  • 3.2 AI creator: The AI creator uses guidance plus Web facts and knowledge to generate new items that fulfill users’ information needs.The paper illustrates this with a personalized music-video creation example.
  • 3.3 Domain Applications of GeneRec: GeneRec is applicable to text, advertisements, video, images, and audio, while micro-video generation requires checks for bias, privacy, authenticity, safety, and realism.The paper describes AI-human collaboration as a possible transition toward AI-driven micro-video creation.
  • 3.4 Potential Research Tasks: Future work includes instruction tuning for LLM-based instructors and control over generator activation and recommendation decisions.These tasks require better intention understanding and decision-making from diverse user behaviors.
  • 3.4 Potential Research Tasks: Personalized item editing should combine user instructions with noisy implicit feedback to capture preferences that users may not articulate.The paper presents this integration as a core research task for AI editors.

4 DISCUSSION

The discussion distinguishes GeneRec from conversational recommendation and traditional content generation, then outlines a future in which generative AI supports increasingly integrated recommendation tasks. It also illustrates editing-task implementations and anticipates unified multimodal generation.

  • 4.1 Comparisons: The discussion compares GeneRec with conversational recommendation and traditional content generation as distinct related tasks.These comparisons frame GeneRec's relationship to existing recommendation and AIGC paradigms.
  • 4.2 A vision for future GeneRec.: GeneRec's envisioned development may begin with AI-assisted repurposing and creation before progressing toward more independent generation in simple scenarios.The roadmap is constrained by the maturity of existing AIGC technologies.
  • 4.2 A vision for future GeneRec.: Figure 5 illustrates AI-editor implementations for thumbnail selection, thumbnail generation, and micro-video clipping.These tasks represent editing-oriented applications within the future GeneRec vision.
  • 4.2 A vision for future GeneRec.: GeneRec is presented as a future information-seeking paradigm that uses user instructions and feedback for retrieval, repurposing, and creation through a unified multimodal model.The paper describes increasing integration among generation tasks as AI technology matures.
  • 4.2 A vision for future GeneRec.: The paper identifies unified multimodal content generation as a feasible and worthwhile technical direction for GeneRec.This direction is intended to integrate multiple generation tasks under one model.

5 FEASIBILITY STUDY

The feasibility study instantiates GeneRec’s AI editor and creator for personalized micro-video selection, generation, revision, and creation. Results show strong personalization for thumbnails, clips, and revision, while personalized creation remains less satisfactory and raises fidelity, copyright, and instruction-complexity concerns.

  • 5.1 AI Editor: RDM achieves the best thumbnail results, while CLIP outperforms Random Frame and Original by using user feedback for personalized selection.RDM can generate thumbnails beyond existing frames, improving alignment with user preference.
  • 5.1 AI Editor: CLIP selects different micro-video clips for users with distinct preferences, and its personalized clips outperform other approaches on Cosine@K and PS@K.The study also compares Random, 1st Clip, and Unclipped baselines.
  • 5.1 AI Editor: Style-transfer cases show high-fidelity, high-quality personalized revisions, but detailed instructions may be costly for more complex editing tasks.The authors identify implicit user feedback as a direction for addressing this burden.
  • 5.1 AI Editor: User-guided revision improves preference alignment, with User_Emb outperforming User_Hist and achieving a smaller FVD score than unconditional revision.The authors attribute User_Emb’s advantage possibly to compressed critical information in pretrained user embeddings, whereas historical features contain noise.
  • 5.2 AI Creator: Personalized micro-video creation improves preference alignment but has worse quality than unconditional creation, including visibly distorted generated content.The authors attribute the weaker results to simple single-turn instructions and limited dataset knowledge, and suggest stronger instruction interfaces, user modeling, and generative models.
  • 5.2 AI Creator: The study emphasizes fidelity checks and highlights unresolved copyright ownership and violation policies for AI-generated or edited micro-videos.It distinguishes platform authorship for fully generated videos from uploader authorship for edits of uploader-created content.

6 RELATED WORK

Related work traces recommender systems from retrieving human- and user-generated content toward generative AI that can complement existing items and capture information needs through multimodal conversations.

  • Traditional recommender systems rank items in a corpus, recommend top-ranked items, and use feedback and context to optimize future rankings.
  • Past work primarily recommends human-generated items, which may fail to satisfy users’ diverse information needs.
  • Generative AI: Content generation evolved from costly expert-created material to user-generated content and increasingly capable generative AI.
  • Generative AI: Generative AI can complement human-generated content, while language models can acquire users’ information needs through multimodal conversations.

7 CONCLUSION AND FUTURE WORK

The paper presents GeneRec as a recommender paradigm combining instruction-guided content generation with retrieval, repurposing, and creation. It identifies user modeling, stronger generation, and trustworthy evaluation as major future directions.

  • GeneRec combines instruction and feedback processing with an AI editor for repurposing and an AI creator for generating new items.
  • The paper highlights fidelity checks and outlines future work on multimodal user modeling, stronger generative modules, and new evaluation standards.
Loading 2304.03516v2…