Research map / Large language models / Retrieval and knowledge

Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models

AAAI 2025 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CIKGRec, which uses gpt-3.5-turbo to infer user interests from behavior, clusters them into a user-interest knowledge graph merged with the CF graph and item KG, and trains a LightGCN-based model with masked reconstruction and cross-domain contrastive learning.

Problem

Knowledge-aware recommenders use only item-side knowledge graphs, while user-side structural knowledge is scarce and coarse; LLM-generated user information is noisy and needs effective transfer into the recommendation graph.

Approach

Use an LLM to infer explicit user interests from historical behaviors, cluster them into interest nodes, and merge them with collaborative and item-side KG into a unified CIKG; denoise via masked graph reconstruction with exponential mask-rate scheduling and transfer knowledge via cross-domain contrastive learning on augmented views.

Key results

• CIKGRec reaches Recall@50 of 0.4642 on DBbook2014 and 0.1791 on Book-Crossing, versus 0.4415 and 0.1653 for the best baselines (5.14% and 8.35% relative gains reported).
• On MovieLens-1M, CIKGRec attains Recall@100 of 0.5718 and NDCG@100 of 0.4609, a modest improvement over the best baselines.
• Ablation shows removing user interest knowledge drops DBbook2014 R@100 from 0.5756 to 0.5673 and Book-Crossing R@100 from 0.2455 to 0.2330.

Measured effects reported in the paper

Surprising findings

• Adding LLM-generated user-side information alone to LightGCN does not consistently help across metrics and datasets, which the authors attribute to noise and lack of a knowledge-transfer module.

Limitations

• LLM-generated user interests may be noisy or hallucinated, and the method relies on a single LLM (gpt-3.5-turbo-0125).
• Relies on text clustering with a hyperparameter kappa that must be tuned per dataset.

Experimental setup

Datasets: DBbook2014 Book-Crossing MovieLens-1M

Benchmarks: DBbook2014 Book-Crossing MovieLens-1M

Metrics: Recall@50 Recall@100 NDCG@50 NDCG@100

Baselines: LightGCN SGL SimGCL CKE CFKG KGAT KGIN KGCL MCCLK KGRec

Builds on: LightGCN SimGCL KGCL GMAE InfoNCE

Closest papers by meaning

• LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.90)
• CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation (AAAI 2025 · similarity 0.90)
• Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.89)
• LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · similarity 0.89)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.89)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.88)