How Alibaba’s New AI-Powered Ad Retrieval Model Redefined E‑Commerce Sponsored Search

The International World Wide Web Conference (WWW) is a premier annual academic event on the future of the Internet. At WWW 2018 in Lyon, Alibaba presented a paper titled “Beyond Keywords and Relevance: A Personalized Ad Retrieval Framework in E‑Commerce Sponsored Search,” which was accepted for oral presentation due to its originality.

Alibaba’s Search Direct Train (also known as Search Ads) is a massive pay‑per‑click advertising platform for Taobao and Tmall sellers, generating a significant portion of the company’s revenue. The 2018 paper disclosed the first public description of a new self‑developed intelligent retrieval model for this platform.

Traditional search‑ad systems rely on advertisers selecting bid keywords. If an advertiser does not purchase the right keywords, relevant ads may never be retrieved, leading to sub‑optimal traffic matching and loss for advertisers, users, and the platform. Moreover, classic models focus solely on relevance, ignoring platform goals such as RPM, CTR, and GMV.

The new model replaces the keyword‑centric approach with a user‑behavior heterogeneous graph that contains three node types—user search signals, ad retrieval keys, and ads—and edges representing intent rewriting and ad recall relationships. By learning edge weights with machine‑learning techniques, the system jointly performs “search intent rewriting” and “ad recall” tasks, automatically constructing rewrite and ad‑recall indexes for online queries. The model also adopts an OCPC (Optimized Cost‑Per‑Click) strategy, allowing automatic bidding without requiring advertisers to purchase keywords.

Deploying this model posed technical challenges: the heterogeneous graph comprises billions of nodes and trillions of edges, making training difficult. Alibaba engineers introduced graph pruning methods to retain important relationships and designed two feature granularities—sparse fine‑grained features for model accuracy and dense coarse‑grained features for coverage and stability.

Online results demonstrated a three‑win outcome: advertisers achieved higher ROI with reduced keyword‑bidding effort, users received more relevant ads improving their experience, and the platform increased revenue while maintaining ecosystem health.