Scholarlink

<p>The rise of large language models (LLMs) has redefined recommender systems, transforming them from static ranking frameworks into dynamic, generative engines that produce contextually rich and explainable recommendations. This review systematically synthesises over 110 peer-reviewed and selected preprint studies (2019-2025). It introduces a reproducible three-family taxonomy comprising Sequence-Driven (SDGS), Knowledge-Driven (KDGS) and Creativity-Driven (CDGS) Generative Systems, each characterised by its dominant generative mechanism. Methodologically, the paper integrates enabling techniques such as prompting, retrieval-augmented generation (RAG), low-rank adaptation (LoRA) and reinforcement learning, linking them to empirical trends across key datasets including Amazon, Goodreads, MovieLens, BookGPT, APIGen, and DimeRec. Across domains such as e-commerce, news, jobs, and entertainment, LLM-based recommenders show improvements in personalisation quality, semantic grounding, dialogue capabilities, and user trust over traditional baselines. This review contextualises LLM-based generative recommendation within the broader data science paradigm, emphasising integrative analytics across text, graph, and user-behaviour modalities. It further identifies challenges of data sparsity, scalability, bias and transparency and maps corresponding technical, ethical, and governance solutions. Theoretical contributions include a unified framework for understanding generative recommendation, while practical insights emphasise fairness-aware, lightweight, and human-centric design. Future work highlights multimodal integration, reinforcement learning from human feedback (RLHF), and cross-domain generalisation as critical steps toward responsible and sustainable deployment of LLM-driven recommender systems.</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

Large language models for generative recommendation: a systematic review of data-centric taxonomy, evaluation, and human-centric analytics

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Takreem Saeed,
  • Bang Wang

摘要

The rise of large language models (LLMs) has redefined recommender systems, transforming them from static ranking frameworks into dynamic, generative engines that produce contextually rich and explainable recommendations. This review systematically synthesises over 110 peer-reviewed and selected preprint studies (2019-2025). It introduces a reproducible three-family taxonomy comprising Sequence-Driven (SDGS), Knowledge-Driven (KDGS) and Creativity-Driven (CDGS) Generative Systems, each characterised by its dominant generative mechanism. Methodologically, the paper integrates enabling techniques such as prompting, retrieval-augmented generation (RAG), low-rank adaptation (LoRA) and reinforcement learning, linking them to empirical trends across key datasets including Amazon, Goodreads, MovieLens, BookGPT, APIGen, and DimeRec. Across domains such as e-commerce, news, jobs, and entertainment, LLM-based recommenders show improvements in personalisation quality, semantic grounding, dialogue capabilities, and user trust over traditional baselines. This review contextualises LLM-based generative recommendation within the broader data science paradigm, emphasising integrative analytics across text, graph, and user-behaviour modalities. It further identifies challenges of data sparsity, scalability, bias and transparency and maps corresponding technical, ethical, and governance solutions. Theoretical contributions include a unified framework for understanding generative recommendation, while practical insights emphasise fairness-aware, lightweight, and human-centric design. Future work highlights multimodal integration, reinforcement learning from human feedback (RLHF), and cross-domain generalisation as critical steps toward responsible and sustainable deployment of LLM-driven recommender systems.