Tubi's Shallow-RHS model cracks cold-start recommendations with asymmetric graph

Tubi, the ad-supported streaming service, faces a classic cold-start problem: newly added content has zero user interaction history, making collaborative filtering and graph-based recommenders ineffective. In production, new content must get a standalone embedding immediately, and device embeddings must support approximate nearest-neighbor retrieval. To solve this, researchers Anh Truong and colleagues propose Shallow-RHS, an asymmetric link-prediction architecture operating on a temporal bipartite device-content graph.

The architecture features a left-hand side (LHS) device tower that leverages temporally valid watch-history message passing to capture collaborative signals, while the right-hand side (RHS) content tower is intentionally shallow—no ID-based embeddings, no content-side subgraphs, no neighbor aggregation. By forcing the RHS encoder to map only intrinsic features (e.g., metadata) into a collaborative-aware embedding space, Shallow-RHS enables implicit graph completion via warm surrogate neighbor retrieval. The same principle is extended to device cold-start using cohort-based embeddings from demographic features. Large-scale online experiments (likely at Tubi) demonstrated consistent relative improvements in content cold-start engagement, promotion speed, impression acquisition, and device cold-start engagement.