Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models (KAR, RecSys 2024)
- Venue: RecSys 2024
- Link: https://dl.acm.org/doi/abs/10.1145/3640457.3688104
- Category: Recommender Engines / LLM-augmented recommendation
- LLMs are most cost-effective in recommendation when used for offline knowledge augmentation, not online generation.
- Combining LLM reasoning factors with standard CF models offers a practical middle ground between traditional recommenders and fully generative (Gen-Rec) systems.
- Hallucination risk, token cost, and feature-refresh pipelines are the main engineering constraints to manage.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Bridging Language and Items for Retrieval and Recommendation (Mar 2024)
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
- DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024
- Original paper: "Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models," RecSys 2024 — ACM DL
Summary
This paper (commonly referred to as KAR, Knowledge Augmentation from Large Language Models) addresses a key limitation of traditional recommendation systems: they learn only from user–item interaction data and struggle with open-world items, long-tail semantics, and cold-start scenarios where collaborative signals are sparse.
KAR augments a conventional recommendation model with knowledge distilled from large language models along two dimensions:
1. Factor extraction — the LLM converts item metadata and user behavior histories into structured, human-interpretable reasoning factors (e.g., item attributes, user preferences). 2. Reasoning — the LLM produces high-level hypotheses about why a user might prefer an item, injecting open-world knowledge beyond what is recorded in logs.
These LLM-generated signals are encoded and fused with the underlying recommendation model through a hybrid-experts style mechanism, so that semantic knowledge complements rather than replaces learned collaborative representations.
Why this design matters
Rather than deploying an LLM inside the online serving path (which is costly and latency-prone), KAR uses the LLM as an offline augmentation pipeline: LLM outputs are precomputed and integrated into the recommender's features. This keeps online inference cheap while still gaining open-world knowledge.
Reported experiments on public benchmarks show consistent ranking improvements over strong baselines, suggesting that LLM-derived factors and reasoning are complementary to ID-based collaborative filtering.