Weibo O-Series Advertising System: Smart Bidding, Intelligent Targeting, and ROI Modeling

Weibo’s O‑Series ads are the mainstream information‑flow advertising format, designed to improve platform revenue by optimizing advertiser ROI through intelligent technology.

The system is divided into three core directions: smart bidding, intelligent targeting, and ROI modeling. Smart bidding uses machine‑learning algorithms to allocate traffic and adjust charges, while intelligent targeting employs deep‑learning to match users and ads. ROI models predict user clicks and conversions (CTR and CVR) to guide bidding and ranking.

Advertising involves a three‑party game among users, advertisers, and the media platform. The platform seeks overall economic gain, advertisers aim for maximum social effect, and the system balances these interests while maximizing platform revenue.

Weibo’s ad products include exposure headers, feed ads, video stories, banner ads, and comment‑stream ads. Billing methods range from traditional CPT, CPM, CPC, CPA, CPS to O‑Series models such as oCPC (conversion‑optimized cost per click) and oCPM (conversion‑optimized cost per mille).

The core technical stack follows a funnel model: coarse targeting, fine ranking, and mechanism strategy, each with specific optimization goals (show rate, click‑through rate, conversion rate).

Smart bidding combines a local greedy algorithm with a global PID control algorithm to balance short‑term and long‑term cost objectives, handling delayed conversion feedback (e.g., app installs) via attribution inference.

Intelligent targeting automatically discovers interested user groups without manual selection, using data‑driven exploration and expansion to maintain ROI while increasing traffic volume.

Modeling relies on DeepFM (factorization machines for second‑order feature interactions plus deep networks for higher‑order interactions) and dual‑tower architectures that separate user and ad representations for efficient training, supplemented by collaborative‑filtering (DeepCF) and clustering for audience segmentation.

Feature engineering incorporates recent user behavior (3‑day windows), keyword searches, user tags, real‑time session features, bias and statistical features, amounting to billions of data points.

A unified offline model framework manages training, evaluation, and deployment of numerous models, providing a consistent feature system and customizable operators for all commercial advertising algorithms at Weibo.