How Kuaishou Leverages Large‑Model AI to Boost Advertising Performance

Overview

Kuaishou has spent the past two years applying large‑model technologies to its advertising scenarios, aiming to better utilize cross‑domain user behavior and external knowledge to enhance ad systems and achieve measurable business gains.

Challenges in Content‑Domain Advertising Estimation

Kuaishou’s platform combines content and e‑commerce, offering various media formats (text, short video, live, product pages, landing pages). User actions are scattered across these formats, making behavior data sparse in any single scenario. Existing recommendation systems rely on ID‑centric pipelines (video ID, item ID, live ID), which are not interoperable, and differing content lifecycles further complicate cross‑domain interest modeling.

Advertising data is especially sparse, but integrating full‑domain behavior—linking natural‑traffic actions with ad interactions—can improve interest modeling. Additionally, recommendation models suffer from “information cocoons,” repeatedly reinforcing the same patterns without external knowledge.

Unified Content Representation Framework COPE

To address ID limitations, Kuaishou introduced a product‑level SPU (Standard Product Unit) ID that aggregates semantically similar items across scenarios, providing a more stable identifier. Building on this, the COPE framework compresses multimodal product content (live, detail pages, short videos) into a robust, compact representation. This unified embedding enriches item features, expands user behavior sequences, and reduces dependence on scenario‑specific data, mitigating sparse‑feature issues.

LLM Knowledge Transfer Framework LEARN

Leveraging the open‑pretrained capabilities of large language models, the LEARN framework injects world knowledge and strong reasoning abilities into Kuaishou’s ad models, helping break the information‑cocoon effect and further enhancing user interest prediction.