ItemRAG: Item-Based Retrieval-Augmented Generation for LLM-Based Recommendation

TL;DR

ItemRAG proposes an item-based retrieval-augmented generation approach for LLM-based recommendations, retrieving relevant items from item–item co-purchase histories to augment the target item's description. For each query item $i$, it builds a retrieval pool $\mathcal{P}(i) = \mathcal{N}(i) \cup \{j : \exists q \in \mathcal{T}(i) \text{ s.t. } j \in \mathcal{N}(q)\}$, where $\mathcal{N}(i)$ are items co-purchased with $i$ and $\mathcal{T}(i)$ are the top-$K$ items similar to $i$ by textual description, and sampling weights $w_{ij} = c_{ij} + \frac{1}{|\mathcal{T}(i)|} \sum_{q \in \mathcal{T}(i)} c_{qj}$; here $c_{ij} = \sum_{u \in \mathcal{U}} \mathbf{1}[\{i,j\} \subseteq \mathcal{M}(u)]$. The retrieved item summaries are appended to the query item description, guiding the LLM to capture co-purchase patterns. Empirically, ItemRAG yields up to 43% improvement in Hit-Ratio@1 and outperforms strong user-based RAG baselines across four Amazon domains under both standard and cold-start item settings.

Abstract

Recently, large language models (LLMs) have been widely used as recommender systems, owing to their strong reasoning capability and their effectiveness in handling cold-start items. To better adapt LLMs for recommendation, retrieval-augmented generation (RAG) has been incorporated. Most existing RAG methods are user-based, retrieving purchase patterns of users similar to the target user and providing them to the LLM. In this work, we propose ItemRAG, an item-based RAG method for LLM-based recommendation that retrieves relevant items (rather than users) from item-item co-purchase histories. ItemRAG helps LLMs capture co-purchase patterns among items, which are beneficial for recommendations. Especially, our retrieval strategy incorporates semantically similar items to better handle cold-start items and uses co-purchase frequencies to improve the relevance of the retrieved items. Through extensive experiments, we demonstrate that ItemRAG consistently (1) improves the zero-shot LLM-based recommender by up to 43% in Hit-Ratio-1 and (2) outperforms user-based RAG baselines under both standard and cold-start item recommendation settings.

Figures (3)

• Figure 1: ItemRAG outperforms the strongest user-based RAG baseline. Across datasets, ItemRAG consistently (1) improves the zero-shot GPT-based recommender and (2) outperforms the strongest user-based RAG baseline, CoRAL wu2024coral.
• Figure 2: An example case of ItemRAG, our item-based RAG method. For retrieving relevant items for item $i$, we first identify items that are co-purchased with (1) item $i$ itself and/or (2) items whose textual descriptions are similar to that of item $i$. Then, we sample a specified number of items from this pool, with selection probabilities proportional to their co-purchase frequencies with item $i$. Subsequently, we prompt an LLM to generate summaries of the sampled items and incorporate these summaries into the final recommendation prompt, guiding the LLM to understand the co-purchase patterns among items.
• Figure 3: (RQ3) Case study. While the naive zero-shot LLM-based recommender fails, augmenting it with co-purchase information retrieved by ItemRAG—information the model explicitly uses—yields an accurate recommendation.