How Kuaishou Uses Large Models to Supercharge Ad Targeting with COPE and LEARN

Overview Over the past two years Kuaishou has explored how large‑model technology can enhance advertising scenarios, focusing on data and model design motivations. The article first outlines challenges in content‑domain ad estimation, then details how multimodal and large‑model techniques improve the ad system by better leveraging global behavior and external knowledge, introducing the COPE product‑content unified representation framework and the LEARN LLM knowledge‑transfer framework.

1. Content‑domain Ad Estimation Challenges and Large‑Model Opportunities Kuaishou’s platform combines content and e‑commerce, offering various media (text, short video, live, product pages) and multiple supply sources (organic creators, advertisers, merchants). User behavior is scattered across these scenarios, making single‑scene data sparse. Existing recommendation systems rely on ID‑centric pipelines (video ID, item ID, live ID), which are not interoperable, hindering cross‑domain interest modeling. Content lifecycles differ (short videos last weeks, live streams hours, products longer), further complicating modeling. Advertising data is especially sparse, but integrating global behavior from natural flow and search can improve interest modeling. Additionally, recommendation loops create information cocoons that are hard to break without external knowledge.

2. Global Behavior Utilization: COPE Unified Product Representation To address ID limitations, Kuaishou built a SPU (Standard Product Unit) system that aggregates same‑product items across scenarios, providing a more stable identifier. However, SPU IDs remain random and lack semantic richness. The COPE framework compresses multimodal product content from live streams, detail pages, short videos, etc., into a robust, compact feature representation. This enhances item‑side features, expands user behavior sequences, and reduces dependence on scenario‑specific interaction data, alleviating insufficient feature learning.

3. External Knowledge Utilization: LEARN LLM Knowledge Transfer By leveraging the open‑pretrained capabilities of large language models, Kuaishou aims to inject world knowledge and strong reasoning/migration abilities to break the information cocoon and improve ad recommendation.