Scholarlink

While Large Language Models (LLMs) demonstrate strong performance in general NLP tasks, their effectiveness in personalized recommendation domains (e.g., news, e-commerce) remains limited. To address this gap and overcome challenges such as noisy interaction data and costly manual curation in existing methods like DPO, we propose CTR-Align, a novel alignment framework that synergizes LLM tuning with CTR prediction models. Our approach introduces two core innovations:(1) A CTR-guided feedback weighting mechanism that dynamically quantifies interaction relevance and user preference strength;(2) A mutually reinforcing paradigm where LLM-generated representations enhance CTR estimation while refined CTR signals guide LLM alignment. Extensive evaluations on MovieLens-1M, Amazon-Toys, and BookCrossing benchmarks show CTR-Align achieves \(>6.9\%\) AUC improvement over state-of-the-art baselines with robust performance.

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

Personalized Recommendations via CTR-Guided LLM Alignment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chao Xu,
  • Xiaowei Gao

摘要

While Large Language Models (LLMs) demonstrate strong performance in general NLP tasks, their effectiveness in personalized recommendation domains (e.g., news, e-commerce) remains limited. To address this gap and overcome challenges such as noisy interaction data and costly manual curation in existing methods like DPO, we propose CTR-Align, a novel alignment framework that synergizes LLM tuning with CTR prediction models. Our approach introduces two core innovations:(1) A CTR-guided feedback weighting mechanism that dynamically quantifies interaction relevance and user preference strength;(2) A mutually reinforcing paradigm where LLM-generated representations enhance CTR estimation while refined CTR signals guide LLM alignment. Extensive evaluations on MovieLens-1M, Amazon-Toys, and BookCrossing benchmarks show CTR-Align achieves \(>6.9\%\) AUC improvement over state-of-the-art baselines with robust performance.