English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

LLMRec: Large Language Models with Graph Augmentation for Recommendation (WSDM 2024)

Forum topic · 小凯 · 2026-07-05

Summary

LLMRec is a WSDM 2024 research paper that applies large language models (LLMs) to collaborative filtering–based recommendation via graph augmentation. Sparse user–item interaction data is a long-standing bottleneck for graph-based recommenders: limited interactions make the user–item bipartite graph noisy and incomplete. LLMRec addresses this by leveraging LLMs' world knowledge to enhance the interaction graph — the LLM infers users' attitudes toward items they have not interacted with, producing LLM-augmented interactions that enrich the graph. The framework incorporates auxiliary prompting strategies and a denoising mechanism to ensure the quality of LLM-inferred feedback, mitigating noise and hallucinations introduced by the augmentation. Experiments on public benchmarks show consistent improvements over competitive baselines, demonstrating that LLM-driven graph augmentation can substantially improve recommendation accuracy while remaining compatible with standard graph neural network recommenders. The paper is part of a growing line of work connecting LLMs with recommender systems, and is catalogued under the Recommender Engines section of the zhichai.net paper list, alongside related entries such as DiffKG and language–item bridging retrieval work.

LLMRec: Large Language Models with Graph Augmentation for Recommendation (WSDM 2024)

Overview

LLMRec is a paper published at WSDM 2024 that integrates Large Language Models (LLMs) with graph augmentation to improve recommendation systems built on collaborative filtering.

  • Paper link: https://dl.acm.org/doi/abs/10.1145/3616855.3635853
  • Venue: WSDM 2024
  • Category: Recommender Engines

    • Motivation

    Graph-based collaborative filtering recommenders are bottlenecked by the sparsity and noise of user–item interaction data:

  • Users interact with only a tiny fraction of the item catalog, so the user–item bipartite graph is highly incomplete.
  • Sparse graphs degrade the quality of learned embeddings and reduce recommendation accuracy.

    • LLMRec's key idea is to use the world knowledge of LLMs to infer users' preferences toward items they have not interacted with, effectively augmenting the interaction graph with high-quality synthetic edges.

    Method

    1. LLM-based augmentation: Auxiliary prompting strategies are used to have the LLM infer user attitudes toward un-interacted items, generating augmented user–item interactions. 2. Denoising: A denoising mechanism (an auxiliary denoising loss) filters unreliable or hallucinated LLM feedback so that noisy augmented edges do not harm training. 3. Recommendation: The enhanced graph is fed into standard graph collaborative filtering models for training and inference.

    Findings

  • Experiments on public benchmark datasets show that LLMRec consistently outperforms competitive recommendation baselines.
  • The results demonstrate that LLM knowledge can be effectively distilled into graph-structured recommenders, improving accuracy even for sparse-interaction users.
  • Careful denoising is essential: naive LLM augmentation can inject noise, and the quality-control step is a core contribution.

    • Context and Related Entries

    This entry is part of a series on LLM-enhanced search and recommendation, cross-referenced with:

  • 360Brew: A Decoder-only Foundation Model for Personalized Ranking
  • Bridging Language and Items for Retrieval and Recommendation
  • DiffKG: Knowledge Graph Diffusion Model for Recommendation (WSDM 2024)
  • Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)

    • Takeaways

  • LLMs can serve as semantic priors to solve the sparsity problem in graph-based recommendation.
  • Augmentation pipelines must include denoising/verification to control hallucination risk.
  • LLM-augmented graphs remain compatible with conventional GNN-based CF architectures, easing industrial adoption.
> Note: Quantitative results should be verified against the original PDF. This summary is based on the paper's public metadata and abstract.

Tags

#llm#recommender-systems#graph-augmentation#collaborative-filtering#wsdm-2024#llmrec#denoising

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208919