Scholarlink

Recommender systems are shaping online user experiences across social and commercial platform. However, they often struggle with cold-start and data sparsity, particularly when user-item interactions are limited or absent. This paper introduces PromptRec-RAG, a modular recommendation framework that integrates large language models (LLMs) with prompt-based conditioning, retrieval-augmented generation (RAG), and synthetic interaction generation. Rather than fine-tuning model weights, PromptRec leverages hard and soft prompts to guide pre-trained LLMs in making accurate predictions under limited data regimes. It improves user and item context by fetching interactions with semantic similarity and generating reasonable feedback through LLMs. We tested the framework for several benchmarks across various cold-start scenarios using the Amazon, Yelp, and MovieLens datasets. PromptRec-RAG outperforms BERT4Rec and DLCRec on NDCG@10 and Recall@10, and audits. Additionally, a 4-way ablation show that its prompts, retrieval, and clean synthetic data together raise NDCG@10 by up to 22%.

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

Prompt-Augmented LLMs with RAG for Addressing Cold-Start and Sparsity in Online Recommender Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sarama Shehmir,
  • Rasha Kashef

摘要

Recommender systems are shaping online user experiences across social and commercial platform. However, they often struggle with cold-start and data sparsity, particularly when user-item interactions are limited or absent. This paper introduces PromptRec-RAG, a modular recommendation framework that integrates large language models (LLMs) with prompt-based conditioning, retrieval-augmented generation (RAG), and synthetic interaction generation. Rather than fine-tuning model weights, PromptRec leverages hard and soft prompts to guide pre-trained LLMs in making accurate predictions under limited data regimes. It improves user and item context by fetching interactions with semantic similarity and generating reasonable feedback through LLMs. We tested the framework for several benchmarks across various cold-start scenarios using the Amazon, Yelp, and MovieLens datasets. PromptRec-RAG outperforms BERT4Rec and DLCRec on NDCG@10 and Recall@10, and audits. Additionally, a 4-way ablation show that its prompts, retrieval, and clean synthetic data together raise NDCG@10 by up to 22%.