Evolution of Weibo Advertising Strategy Engineering Architecture

The talk introduces the theme "Weibo Advertising Strategy Engineering Architecture Evolution" and outlines how the advertising system has progressed from a zero‑to‑one stage to a mature N‑stage platform that supports strategy, algorithm, and model iteration.

1. Overview of Advertising Styles and Scenarios – Weibo’s commercial flow consists of a "one screen, four major streams" model covering relationship, hot, comment, and search streams, with a backend for ad placement.

2. Advertising Participants – Advertisers are classified as KA (large) or small‑mid, with common billing methods such as CPE, CPM, CPD, and the industry‑wide OCPX model.

3. Core Problems of Computational Advertising – The system must balance the interests of the platform, users, and advertisers to maximize overall benefit.

4. Advertising Delivery Process – A typical workflow includes creating a promotion plan, audience targeting, budget setting, creative design, launch, performance monitoring, and subsequent marketing decisions.

5. Architecture Evolution – The architecture evolved from early banner ads to version 4.0 in 2017, introducing a layered design that emphasizes high availability, high concurrency, and support for rapid algorithmic iteration.

6. Support for Advertising Growth Transformation – The system shifts from coarse‑grained growth (expanding traffic, advertisers, and budget) to fine‑grained growth by enhancing model‑driven strategies, including deep‑learning‑based DNN vector triggers.

7. DNN Vector Trigger Model – A dual‑tower architecture generates user and ad embeddings via three‑layer neural networks, using cosine similarity and sigmoid for relevance scoring.

8. Trigger Engineering Architecture – Five‑path recall (dual‑tower, content, user profile, precise audience, social graph) is merged by a Mixer and refined by quality estimation before ranking.

9. Lean‑Driven "Two‑Wing" Plan – The Faraday experimental platform and Faraday lean insight provide a systematic way to run online strategy experiments and monitor their impact, employing orthogonal layered experiments and traffic bucketing.

10. Experimentation and Evaluation – Real‑time effect evaluation (≈5‑minute latency) and offline cross‑day analysis ensure reliable metric tracking for up to hundreds of millions of business indicators.

Overall, the presentation demonstrates how Weibo’s advertising system integrates sophisticated engineering, algorithmic models, and lean experimentation to achieve scalable, high‑performance ad delivery.