Exploring and Applying Large Language Models in Recommendation Systems

Overview – Wang Yichao from Huawei Noah's Ark Lab delivered a presentation on the exploration and application of large language models (LLMs) in recommendation systems, discussing data, model, and workflow aspects, and introducing two key Huawei projects while addressing user reasoning knowledge construction, feature crossing, and online service processes.

Background and Problem – Traditional recommendation systems are closed, relying on domain‑specific logs and lacking external world knowledge. LLMs bring rich factual and commonsense knowledge plus logical reasoning, which can supplement testing, scoring, and workflow control, either via fine‑tuning during training or as inference engines.

LLM4Rec – KAR (Knowledge‑Assisted Recommendation) – KAR uses LLMs to generate open‑world knowledge for user preference and item factual reasoning, transforms this knowledge into low‑dimensional dense vectors via a multi‑expert network, and integrates it with traditional recommenders to improve AUC by ~1% while keeping inference latency comparable to baseline models.

LLM4Rec – Uni‑CTR (Unified Cross‑Domain Recommendation) – Uni‑CTR leverages LLMs to build a multi‑scene recommendation foundation. It converts tabular features into natural‑language prompts, feeds them to a 24‑layer Transformer (SharedBert), and employs Leader and Backbone networks to capture scene‑specific and shared representations, achieving zero‑shot capability and balanced performance across Fashion, Music Instruments, and Gift Cards datasets.

Experimental Results – Both KAR and Uni‑CTR were deployed in Huawei’s app market, music, and advertising scenarios. Experiments on Amazon Review datasets show significant AUC improvements, especially in sparse domains, and demonstrate that combining user‑preference and item‑factual knowledge yields the best gains.

Challenges and Outlook – Remaining challenges include joint modeling of collaborative and semantic signals, efficient handling of long textual inputs and ID encoding, and real‑time integration of dynamic data. Future work will focus on opening recommendation systems to world knowledge, shifting from discriminative to generative models, and moving toward end‑to‑end unified recommendation pipelines.