Scholarlink

Personalized book recommendation systems play a crucial role in enhancing user experience by guiding readers toward literature that aligns with their unique interests. While effective, traditional collaborative filtering (CF) techniques often suffer from sparsity and limited interpretability. Recent advances in natural language processing have introduced retrieval-augmented generation (RAG) models, which enable the generation of textual content for context-aware recommendation. This study proposes a novel hybrid framework integrating CF with RAG to generate accurate and semantically rich book recommendations. The CF component models latent user-item interactions using matrix factorization, while the RAG component retrieves and conditions on relevant book descriptions and reviews. We evaluate our approach on two benchmark datasets: Goodbooks-10K and Book-Crossing. Experimental results show that our method achieves improvements of 4–7% in HR@10 and NDCG@10 over strong CF and language model baselines, while also producing more coherent recommendation texts as reflected in higher BLEU-2 scores. These findings highlight the effectiveness of combining structured and unstructured information in personalized recommendation tasks.

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

Retrieval-Augmented Generation with Collaborative Filtering for Personalized Book Recommendations

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Luong Vuong Nguyen

摘要

Personalized book recommendation systems play a crucial role in enhancing user experience by guiding readers toward literature that aligns with their unique interests. While effective, traditional collaborative filtering (CF) techniques often suffer from sparsity and limited interpretability. Recent advances in natural language processing have introduced retrieval-augmented generation (RAG) models, which enable the generation of textual content for context-aware recommendation. This study proposes a novel hybrid framework integrating CF with RAG to generate accurate and semantically rich book recommendations. The CF component models latent user-item interactions using matrix factorization, while the RAG component retrieves and conditions on relevant book descriptions and reviews. We evaluate our approach on two benchmark datasets: Goodbooks-10K and Book-Crossing. Experimental results show that our method achieves improvements of 4–7% in HR@10 and NDCG@10 over strong CF and language model baselines, while also producing more coherent recommendation texts as reflected in higher BLEU-2 scores. These findings highlight the effectiveness of combining structured and unstructured information in personalized recommendation tasks.