Scholarlink

<p>Social recommendation enhances recommendation quality by incorporating social relationships into user preference learning. Graph-based deep learning methods have achieved strong performance in social recommendation, though they are mostly ID-based. However, existing ID-based methods mainly rely on user interaction behaviors, making it difficult to leverage rich textual information and limiting generalization capability (Challenge 1). Large Language Models (LLMs), possessing powerful capabilities in comprehending natural language, offer a promising solution when combined with graph-based models. Nevertheless, LLMs face two major challenges in understanding graph-structured data: (1) converting graph data into natural language produces lengthy descriptions, making it difficult for LLMs to extract user interaction information (Challenge 2); and (2) embedding interaction features into textual prompts to align with the LLM input space may disrupt the model’s interpretation of the original recommendation content (Challenge 3). To overcome these issues, we introduce a novel social recommendation model, LLM-SocRec. The model first constructs a user–item interaction graph along with a user–user social graph to derive embeddings for both users and items. These embeddings are transformed into low-rank collaborative-aware weights and integrated into LLM parameters. This design eliminates the need to convert graph information into verbose text while preserving the LLM’s ability to capture the semantics of original text. We additionally incorporate Direct Preference Optimization (DPO) to better align with user preferences and boost recommendation effectiveness. Experimental results show that LLM-SocRec outperforms existing baselines across three datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-SocRec: Enhancing Graph-based Social Recommendation via Collaborative Large Language Models

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Ziyu Li,
  • Zhijie Tan,
  • Weiping Li,
  • Tong Mo

摘要

Social recommendation enhances recommendation quality by incorporating social relationships into user preference learning. Graph-based deep learning methods have achieved strong performance in social recommendation, though they are mostly ID-based. However, existing ID-based methods mainly rely on user interaction behaviors, making it difficult to leverage rich textual information and limiting generalization capability (Challenge 1). Large Language Models (LLMs), possessing powerful capabilities in comprehending natural language, offer a promising solution when combined with graph-based models. Nevertheless, LLMs face two major challenges in understanding graph-structured data: (1) converting graph data into natural language produces lengthy descriptions, making it difficult for LLMs to extract user interaction information (Challenge 2); and (2) embedding interaction features into textual prompts to align with the LLM input space may disrupt the model’s interpretation of the original recommendation content (Challenge 3). To overcome these issues, we introduce a novel social recommendation model, LLM-SocRec. The model first constructs a user–item interaction graph along with a user–user social graph to derive embeddings for both users and items. These embeddings are transformed into low-rank collaborative-aware weights and integrated into LLM parameters. This design eliminates the need to convert graph information into verbose text while preserving the LLM’s ability to capture the semantics of original text. We additionally incorporate Direct Preference Optimization (DPO) to better align with user preferences and boost recommendation effectiveness. Experimental results show that LLM-SocRec outperforms existing baselines across three datasets.