Hulu’s AI Secrets: Binge Ad Prediction and Real‑Time Allocation
On November 7, 2020, Hulu’s Shulei presented how the company’s video advertising platform employs AI-driven algorithms—including binge‑watch prediction, inventory forecasting with Prophet, and a PID‑controller based real‑time allocation system—to enhance targeting, optimize revenue, and improve user experience.
Introduction
On November 7, 2020, Tsinghua University’s Computer Science Department, AI Time, and Xuetang Online co‑hosted a technical sharing event where Hulu’s Shulei presented the algorithms used in Hulu’s video advertising system.
Hulu Advertising Business Overview
Hulu’s ad revenue primarily comes from brand advertising billed per thousand impressions, focusing on reach, frequency, and brand awareness. The ad inventory consists of contract ads—pre‑negotiated guaranteed impressions—and auction ads, which are sold via SSP platforms in a private marketplace.
Ad Lifecycle and Core Processes
A typical ad lifecycle includes two major phases: sales & planning, and delivery & settlement. These comprise four key processes:
Demand Capture : Sales teams gather advertiser requirements and use targeting algorithms to define precise audience segments.
Order Management : Systems confirm exposure volume, pricing, and schedule, employing inventory forecasting, reach and frequency prediction, and dynamic pricing algorithms.
Online Delivery : The real‑time bidding platform allocates ad slots using online inventory allocation algorithms.
Monitoring & Settlement : Real‑time fraud detection and attribution analysis ensure accurate measurement and billing.
Algorithm Highlights
The talk focused on three representative algorithms:
Binge‑Ad Prediction : Designed for Hulu’s new “Binge Ad” format, this model identifies users likely to watch three or more consecutive episodes. It evolved from offline user‑based predictions to near‑line and real‑time features, and finally to a CNN that models recent viewing sequences.
Inventory Forecasting : Utilizes Facebook’s Prophet time‑series model, which decomposes traffic into trend, yearly, weekly, and holiday components. Two practical improvements were introduced to address real‑world challenges.
Online Inventory Allocation : Implements a PID‑controller‑based dynamic allocation algorithm, treating the problem as a bipartite graph matching. The controller continuously adjusts coefficients for each ad request, supporting both contract and auction ads to boost revenue and relevance.
Conclusion
The presented algorithms represent only a fraction of Hulu’s advertising technology stack. Ongoing research aims to further enhance ad products and user experience.
Hulu Beijing
Follow Hulu's official WeChat account for the latest company updates and recruitment information.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.