Knowledge Graph Enhanced Language Agents for Recommendation

TL;DR

The paper tackles the limitation of language-model–driven recommender simulations that rely on vague user profiles by introducing Knowledge Graph Enhanced Autonomous Language Agents (KGLA). It unifies three modules—Path Extraction, Path Translation, and Path Incorporation—to convert knowledge-graph paths between known users and items into informative, text-based rationales that enrich agent memories and guide both simulation and ranking. Empirical results on CDs, Clothing, and Beauty show large $NDCG@1$ gains over baselines (up to about 95% relative improvements) and demonstrate the method's effectiveness across different LLMs, with stronger performance when combining $2$-hop and $3$-hop KG information. The approach also provides explainable recommendations and is particularly advantageous in low-interaction or cold-start scenarios, offering a pathway toward more interactive and memory-grounded recommender systems.

Abstract

Language agents have recently been used to simulate human behavior and user-item interactions for recommendation systems. However, current language agent simulations do not understand the relationships between users and items, leading to inaccurate user profiles and ineffective recommendations. In this work, we explore the utility of Knowledge Graphs (KGs), which contain extensive and reliable relationships between users and items, for recommendation. Our key insight is that the paths in a KG can capture complex relationships between users and items, eliciting the underlying reasons for user preferences and enriching user profiles. Leveraging this insight, we propose Knowledge Graph Enhanced Language Agents(KGLA), a framework that unifies language agents and KG for recommendation systems. In the simulated recommendation scenario, we position the user and item within the KG and integrate KG paths as natural language descriptions into the simulation. This allows language agents to interact with each other and discover sufficient rationale behind their interactions, making the simulation more accurate and aligned with real-world cases, thus improving recommendation performance. Our experimental results show that KGLA significantly improves recommendation performance (with a 33%-95% boost in NDCG@1 among three widely used benchmarks) compared to the previous best baseline method.

Figures (5)

• Figure 1: Examples of user agent memory generated (a) without KG and (b) with KG. The user agent memory generated without KG only contains some general descriptions, while the memory generated with KG includes more specific terms (highlighted in red), demonstrating more precise user preferences.
• Figure 2: The framework of our proposed KG-enhanced Agent Simulation for recommendation. Given $(u, i^+, i^-)$, our framework can retrieve and translate KG information which would be incorporated into the Simulation, guiding LLM agents to analyze the possible reasons for user choices based on KG, summarize the user's precise preferences, and update the user agent's memory.
• Figure 3: Experimental results across different numbers of users, with NDCG@1, NDCG@5, and NDCG@10 reported separately.
• Figure 4: A case study for KG-enhanced simulation: Features in bold font such as "sensual" and "sultry" are summarized by the User Agents as the preferred features while light-colored features such as "pray" and "fallen" are summarized by User Agents as the non-informative features.
• Figure 5: A case study for KG-enhanced ranking: Taking two item candidates as examples, Candidate 1 aligns more closely with the user agent's memory, enabling the user agent to make the correct choice during the ranking stage.