Adaptive Candidate Retrieval with Dynamic Knowledge Graph Construction for Cold-Start Recommendation

TL;DR

ColdRAG tackles item cold-start in recommender systems by dynamically building a knowledge graph from raw item metadata and using LLM-guided multi-hop reasoning to adaptively retrieve and rank candidates without pre-filtered lists. The framework combines Item Profile Generation and Knowledge Graph Generation to form a semantically rich graph, followed by Adaptive Candidate Retrieval over KG and Retrieval-augmented Generation for final recommendations. Empirical results across multiple Amazon domains show consistent gains over both training-based and training-free baselines, with strong stability and reduced hallucination due to graph-grounded reasoning. The work demonstrates practical impact by enabling robust zero-shot recommendations in dynamic catalogs while highlighting considerations around compute, latency, and hyperparameter sensitivity.

Abstract

The cold-start problem remains a critical challenge in real-world recommender systems, as new items with limited interaction data or insufficient information are frequently introduced. Despite recent advances leveraging external knowledge such as knowledge graphs (KGs) and large language models (LLMs), recommender systems still face challenges in practical environments. Static KGs are expensive to construct and quickly become outdated, while LLM-based methods depend on pre-filtered candidate lists due to limited context windows. To address these limitations, we propose ColdRAG, a retrieval-augmented framework that dynamically constructs a knowledge graph from raw metadata, extracts entities and relations to construct an updatable structure, and introduces LLM-guided multi-hop reasoning at inference time to retrieve and rank candidates without relying on pre-filtered lists. Experiments across multiple benchmarks show that ColdRAG consistently outperforms strong seven baselines.

Figures (10)

• Figure 1: Overview of the proposed ColdRAG framework. Given item metadata, an LLM generates item profiles, from which structured entities and relations form a knowledge graph dynamically. During inference, ColdRAG performs query-aware multi-hop reasoning over KG to adaptively retrieve candidate items and context, then composes prompts to generate recommendations.
• Figure 2: Performance comparison of ColdRAG variants across three domains using GPT, showing that both core modules (G and R) add performance gains.
• Figure 3: (a) Recall@10 box plots over five runs for training-free baselines. (b) Out-of-domain generation rates, both evaluated on the Games dataset with GPT.
• Figure 4: Illustration of ColdRAG’s adaptability to item cold-start scenario.
• Figure 5: (a) Pie chart of the distribution of entity (node) types in the KG. (b) Heatmap of the number of relations (edges) between the entity types.