Algorithmic Practices in Hulu's Video Advertising System

Hulu, a leading US video‑on‑demand and live‑streaming service, relies heavily on advertising as a primary revenue source, focusing mainly on brand‑type video ads that are sold through guaranteed‑delivery contracts.

The advertising ecosystem involves three core participants—advertisers, users, and the media platform—each with distinct optimization goals such as ad delivery efficiency, return‑on‑ad‑spend (ROAS), user experience, and overall ad revenue maximization.

Key algorithmic problems addressed include:

Ad targeting: using machine‑learning models (e.g., XGBoost, DNN) for user‑profile completion, look‑alike modeling, and context‑aware targeting via image and audio recognition.

Inventory prediction: forecasting future ad‑slot supply with time‑series models such as ARIMA, Prophet, and LSTM, accounting for user growth, seasonality, and holidays.

Inventory allocation (traffic matching): formulating a bipartite‑graph matching problem between supply (user traffic) and demand (advertiser orders), solving it offline with Lagrangian duality and online PID control for real‑time adjustments.

Conversion‑rate optimization: building DIN + FM models for CVR estimation, incorporating user, ad, and user‑ad interaction features, and applying causal inference (doubly robust estimator) to obtain unbiased uplift metrics.

Programmatic real‑time bidding and pricing strategy: using bandit and reinforcement‑learning methods for ad cold‑start and dynamic pricing.

Shared‑account handling: detecting multiple virtual users behind a single account, labeling their behavior patterns, and personalizing ad delivery accordingly.

Image‑based contextual targeting leverages pretrained networks (Inception V3, VGG) fine‑tuned on Hulu‑specific taxonomy to identify objects, scenes, and moods, ensuring ads align with video content and avoid inappropriate placements.

Offline traffic matching is modeled as a bipartite graph where supply nodes represent user traffic with attribute tags and demand nodes represent advertiser orders with targeting constraints; edge weights are optimized to satisfy both inventory limits and order fulfillment requirements.

Online adjustments employ PID controllers to align actual delivery speed with planned schedules, maintaining smooth pacing and handling deviations in real time.

The article concludes by highlighting ongoing research challenges such as improving causal‑inference‑based uplift modeling, refining shared‑account detection, and exploring more advanced optimization techniques for large‑scale ad systems.