Scholarlink

<p>Dual-target cross-domain recommendation aims at improving recommendation accuracy of both domains. Most studies transfer user behavioral knowledge across domains under the multi-task learning paradigm, which struggles to combat domain heterogeneity and impairs knowledge transfer effectiveness. To address this challenge, we propose LLMeta, a large language models (LLM)-enhanced meta-learning method that learns and then flexibly adapts abundant domain-generalized knowledge to individual domains. First, we develop an LLM-enhanced unified graph model that utilizes LLMs to extract cross-domain common profile knowledge and then transfers such knowledge into the unified graph model. Second, we employ model-agnostic meta-learning to meta-train the parameters in the unified graph model to ensure these meta-parameters encode domain-generalized knowledge. Finally, we develop a novel domain-specific adaptation module to comprehensively adapt both the meta-trained underlying embeddings and the global collaborative signals hidden in meta-trained final representations. Experiments on Elec &amp; Cell, Sports &amp; Cloth and Sports &amp; Cell datasets from Amazon validate that our LLMeta significantly surpasses competitive baselines, with the average accuracy improvements on HR@5 of 3.20%, 6.54% and 4.78% in two domains of these three datasets, respectively.</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

Learning and adapting domain-generalized knowledge for dual-target cross-domain recommendation via LLM-enhanced meta-learning

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Mingxin Gan,
  • Jieyu Ren

摘要

Dual-target cross-domain recommendation aims at improving recommendation accuracy of both domains. Most studies transfer user behavioral knowledge across domains under the multi-task learning paradigm, which struggles to combat domain heterogeneity and impairs knowledge transfer effectiveness. To address this challenge, we propose LLMeta, a large language models (LLM)-enhanced meta-learning method that learns and then flexibly adapts abundant domain-generalized knowledge to individual domains. First, we develop an LLM-enhanced unified graph model that utilizes LLMs to extract cross-domain common profile knowledge and then transfers such knowledge into the unified graph model. Second, we employ model-agnostic meta-learning to meta-train the parameters in the unified graph model to ensure these meta-parameters encode domain-generalized knowledge. Finally, we develop a novel domain-specific adaptation module to comprehensively adapt both the meta-trained underlying embeddings and the global collaborative signals hidden in meta-trained final representations. Experiments on Elec & Cell, Sports & Cloth and Sports & Cell datasets from Amazon validate that our LLMeta significantly surpasses competitive baselines, with the average accuracy improvements on HR@5 of 3.20%, 6.54% and 4.78% in two domains of these three datasets, respectively.