Scholarlink

<p>Recommender systems significantly enhance user experience by generating personalized suggestions through the analysis of user behavior. In recommendation algorithms, collaborative filtering and deep learning approaches utilizing Graph Neural Networks (GNNs) and Transformers have achieved strong performance. However, most recommendation algorithms primarily focus on the accuracy of the results, without explicitly revealing the reasoning process behind the recommendations—that is, the rationale for the recommendations. To bridge this gap, explainable recommendation algorithms have been proposed to provide users with an explicit reasoning process behind the recommendations. However, existing explainable recommendation algorithms face three major challenges: (1) a focus on fitting user reviews rather than genuinely generating recommendation rationales, (2) limited generalization ability in zero-shot scenarios, and (3) the scarcity of high-quality explanation data. To this end, we propose STLLM-Rec, a <Emphasis Type="Underline">S</Emphasis>elf-<Emphasis Type="Underline">T</Emphasis>raining framework designed to enhance the reasoning capability of <Emphasis Type="Underline">L</Emphasis>arge <Emphasis Type="Underline">L</Emphasis>anguage <Emphasis Type="Underline">M</Emphasis>odels (LLMs) for <Emphasis Type="Underline">Rec</Emphasis>ommendation tasks. This approach aims to leverage the powerful reasoning and generalization capabilities of LLMs to generate explicit recommendation rationales (addressing Challenge 1) and enhance the generalization ability of recommendation systems in zero-shot scenarios (addressing Challenge 2). In addition, it tackles the challenge of scarce high-quality explanation data (Challenge 3) by introducing self-training techniques from LLMs and incorporating fine-grained reward signals to guide the reasoning and training processes. Extensive experiments demonstrate that STLLM-Rec achieves both high recommendation accuracy and strong explainability.</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

STLLM-Rec: enhancing explainable recommendation via self-training LLMs

  • 期刊论文

  • 发表日期：

    2026年1月9日

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

摘要

Recommender systems significantly enhance user experience by generating personalized suggestions through the analysis of user behavior. In recommendation algorithms, collaborative filtering and deep learning approaches utilizing Graph Neural Networks (GNNs) and Transformers have achieved strong performance. However, most recommendation algorithms primarily focus on the accuracy of the results, without explicitly revealing the reasoning process behind the recommendations—that is, the rationale for the recommendations. To bridge this gap, explainable recommendation algorithms have been proposed to provide users with an explicit reasoning process behind the recommendations. However, existing explainable recommendation algorithms face three major challenges: (1) a focus on fitting user reviews rather than genuinely generating recommendation rationales, (2) limited generalization ability in zero-shot scenarios, and (3) the scarcity of high-quality explanation data. To this end, we propose STLLM-Rec, a Self-Training framework designed to enhance the reasoning capability of Large Language Models (LLMs) for Recommendation tasks. This approach aims to leverage the powerful reasoning and generalization capabilities of LLMs to generate explicit recommendation rationales (addressing Challenge 1) and enhance the generalization ability of recommendation systems in zero-shot scenarios (addressing Challenge 2). In addition, it tackles the challenge of scarce high-quality explanation data (Challenge 3) by introducing self-training techniques from LLMs and incorporating fine-grained reward signals to guide the reasoning and training processes. Extensive experiments demonstrate that STLLM-Rec achieves both high recommendation accuracy and strong explainability.