Hulu’s AI Innovations: Graph Neural Networks, Ad Targeting & Content Embeddings

Introduction

At the beginning of 2020, Hulu partnered with People’s Posts & Telecommunications Press to publish "Hundred‑Faced Deep Learning," which received wide acclaim. Earlier, the 2018 "Hundred‑Faced Machine Learning" series became a staple for students and AI practitioners. Hulu, a Disney‑owned US streaming service, operates its only overseas office in Beijing, focusing on personalized recommendation, content discovery, algorithmic operations, video codec optimization, and personalized ad delivery. To thank the community for supporting the "Hundred‑Faced" series, Hulu, the press, and DataFun organized this offline technical salon to share algorithm projects and engineering practices.

Share Topic 1: Algorithms for Large‑Scale Graph Neural Network Computation

Hulu research engineer Xiaoran presented optimization ideas for large‑scale graph neural networks. The talk began with a review of the basic message‑passing framework, where each propagation iteration computes messages from source to target nodes, aggregates received messages, and updates node states. Nine optimization strategies were discussed, covering dimension decomposition, node partitioning, sparse graph computation, temporary memory modules, and distillation methods.

Share Topic 2: Multi‑Factor Placement Algorithms in Video Advertising Systems

Chunyang from the advertising algorithm team introduced the multi‑factor placement algorithm for Hulu’s video ad system. The presentation covered three parts: an overview of Hulu’s main ad formats and research problems, design of a placement algorithm considering volume control, pacing, frequency, priority, platform revenue, and user experience, and practical experiences with the algorithm evaluation platform.

Share Topic 3: Content Retrieving and Ranking in Hulu

Jiarui from the recommendation algorithm group described Hulu’s recommendation scenarios, including content recall, category recall, related recall, and list recall. He highlighted the system’s multi‑scene, multi‑model nature, small post‑selection sets, core optimization metrics, and two distribution modes for content discovery and continuous watching. Key techniques such as next‑item prediction, CTR prediction, multi‑task learning, and relevance merging were presented.

Share Topic 4: Hulu Search Engine Technology Practice

Min from the search team explained that accurate understanding of user intent is fundamental to search. Personalized search combines query terms with user behavior. The team revamped a classic Siamese network for content retrieval and employed a Wide&Deep model at the ranking stage, integrating user, query, and document features. Data collection spans offline, near‑line, and real‑time pipelines across homepage, search page, and other surfaces.

Share Topic 5: Machine Learning in Video Codec and Streaming

Wenhao from the video algorithm team discussed how Hulu leverages machine‑learning models across video encoding and streaming pipelines to improve visual quality for its diverse content library and wide range of devices.

Share Topic 6: Content Embeddings in Hulu – Generation & Application

Yunsheng from the video content understanding group described Hulu’s rich metadata (titles, descriptions, directors, actors, genres, release dates, awards, keywords) and video data for each show. Content embedding transforms this multimodal information into low‑dimensional vectors, enabling similarity computation, serving as features for recommendation models, aiding cold‑start, and supporting collection modeling. The team employed Graph Embedding for tag data, Metadata‑BERT for textual data, and Two‑Level‑BERT for video data, producing tag‑based, textual, and visual/audio embeddings that are applied across various business scenarios.

-END-