Skip to main content
Back

Tubi's Shallow-RHS Fixes Blind Recommendations for 40,000 Cold-Start Titles

Tubi's Shallow-RHS architecture solves cold-start recommendation by training a graph-aware content encoder that works without any interaction history, boosting engagement for new content and new users at scale.

Paper, Full Access
  • Tubi's Shallow-RHS is a new asymmetric two-tower recommendation architecture that solves cold-start by training a content encoder on intrinsic features only, with no interaction history required.
  • The content (RHS) tower deliberately uses no ID embeddings, no graph neighbors, and no interaction signals -- forcing it to map metadata into a collaborative-filtering-aware space.
  • The device (LHS) tower runs full GNN message passing over temporally valid watch history to capture rich collaborative signals.
  • New titles get an embedding immediately upon ingestion and find warm surrogate neighbors via ANN retrieval -- implicit graph completion with no special inference path.
  • Device cold-start is handled via cohort-based embeddings built from demographic features, giving new users a starting point without any watch history.
  • Large-scale A/B tests on Tubi's live system showed consistent improvements in cold-start engagement, promotion speed, and impression acquisition; no code is publicly released.

Every streaming platform has a dirty secret: the moment new content lands in the catalog, the recommendation system goes blind. A new paper from Tubi's engineering team describes how they tackled this head-on, building a production system that can embed brand-new titles the instant they are ingested -- no watch history required -- and still surface them to the right viewers through collaborative-filtering-aware retrieval.

The stakes are real. Tubi exceeded 100 million monthly active users and its content library contains nearly 300,000 TV episodes and movies -- approximately 40,000 distinct titles. Free ad-supported streaming services recorded the fastest catalogue growth in 2025, outpacing subscription video-on-demand platforms worldwide, with Tubi UK alone growing its catalogue 99% in a single year. When you are adding titles that fast, cold-start is not an edge case -- it is the norm.

The Wall Every Recommender Hits

When items are added to the catalogue, the item cold-start problem occurs because they have no or very few interactions. This is particularly problematic for collaborative filtering algorithms, which generate recommendations based on the item's interactions -- a pure collaborative algorithm simply cannot recommend an item if no interactions are available.

The standard fix is a two-tower model: one neural network encodes the user (or device), another encodes the item, and you retrieve candidates by finding the nearest neighbors in the shared embedding space. The Two-Tower model is a foundational architecture for large-scale recommendation systems, built to efficiently retrieve relevant items from massive catalogs, enabling fast candidate generation via approximate nearest neighbor (ANN) search -- critical for real-time personalization. The problem is that most two-tower implementations still rely on learned ID embeddings for items, which means a brand-new title has no embedding at all until it accumulates interactions.

Tubi's production system adds a further constraint that makes this worse. New content must be assigned a standalone embedding immediately upon ingestion, and device (user) embeddings must be compatible with ANN retrieval. You cannot retrain the model every time a new movie lands. The system has to generalize.