Applying Machine Learning to Advertising‑Based Video‑On‑Demand (AVOD) at Tubi

In this series blog, Tubi describes how machine learning is used across its advertising‑based video‑on‑demand (AVOD) service, addressing challenges in recommendation, content understanding, and advertising by heavily employing PyTorch for large‑scale GPU training.

Tubi is a leading free streaming platform with over 33 million monthly active users and more than 2.5 billion hours of watch time last year, offering movies, TV shows, and live news across dozens of devices.

The AVOD model relies on three interrelated pillars: Content (the video catalog), Audience (viewers), and Advertising (ads shown to viewers). Maximizing satisfaction across these pillars creates a virtuous cycle that drives revenue and content acquisition.

Machine learning is woven into this ecosystem in three key areas: Content Understanding , Advertising Technology , and Recommendation Systems . The article outlines each area and promises deeper future posts.

Content Understanding (the "Spock" project) uses embeddings generated from first‑ and third‑party data, leveraging NLP techniques from word2vec and fastText to modern Transformer models such as ELMo, BERT, and BigBird. PyTorch models handle cold‑start scenarios, mapping high‑dimensional embeddings to Tubi’s collaborative‑filtering space, enabling predictions of new titles’ value.

Advertising Technology focuses on delivering pleasant ad experiences by targeting audiences, controlling ad placement timing and frequency, and optimizing revenue through advanced frequency‑management (AFM) models built with PyTorch that detect and limit brand exposure using computer‑vision techniques.

Recommendation Systems aim to surface the most relevant videos, handling challenges such as massive user scale, short‑lived news content, and a growing catalog. Tubi’s stack includes Spark, MLeap, MLflow, and Databricks, with ongoing research into neural collaborative filtering using PyTorch and rich data from the content‑understanding pipeline.

The underlying technical stack relies on AWS, Azure, Databricks, Spark, and a suite of ML libraries (PyTorch, XGBoost, etc.) to process billions of records, support low‑latency real‑time inference, and enable rapid experimentation and deployment.

In summary, the article provides an overview of Tubi’s AVOD business model, the three ML‑driven pillars supporting it, and the infrastructure that powers their machine‑learning workflows, with promises of deeper dives into each pillar in upcoming posts.