How Kuaishou Uses Large Models to Boost Ad Performance with COPE and LEARN

Overview

In the past two years Kuaishou has explored the use of large‑model technology in advertising scenarios, presenting the motivations and design considerations behind its data and model innovations.

Challenges in Content‑Domain Ad Estimation

Kuaishou’s platform combines content and e‑commerce, offering multiple media formats (text‑image, short video, live, product pages). User behavior is scattered across these contexts, making single‑scene data sparse. ID‑centric recommendation systems (video ID, item ID, live ID) hinder cross‑domain interest modeling, and differing content lifecycles add complexity.

Opportunity: Full‑Domain Behavior Utilization

By unifying behavior across recommendation and search streams, Kuaishou aims to build richer user interest models. Incorporating external knowledge via large language models (LLM) can break the “information cocoon” that limits purely behavior‑driven models.

COPE: Unified Product Content Representation Framework

Traditional item ID embeddings lack semantic meaning and struggle with new items. Kuaishou introduced an SPU‑ID system to aggregate same‑product items, but it remains random and behavior‑dependent. COPE compresses multimodal product content from live, detail pages, and short videos into a robust, compact feature vector, enhancing item representations and reducing reliance on sparse scene‑specific behavior.

LEARN: LLM Knowledge Transfer Framework

Leveraging pretrained LLMs’ world knowledge and reasoning ability, the LEARN framework transfers external knowledge into advertising models, helping to overcome data sparsity and break the recommendation loop.