Cracking the Cold-Start: Tubi's Revolutionary Approach

The cold-start problem has long plagued recommendation systems. New content struggles to find an audience because it lacks interaction history. Tubi, the streaming service, is tackling this issue head-on with a fresh approach that could set a new standard.

The Cold-Start Challenge

Recommendation systems typically rely on user interaction data to suggest content. That’s why the cold-start problem is so tricky. New material doesn’t have this data. Tubi’s solution? Treating it as an inductive graph-completion problem. They focus on a temporal bipartite device-content graph, aiming to generate immediate and effective content embeddings. This is essential for quick retrieval and user engagement.

Shallow-RHS: A Unique Architecture

Tubi introduces Shallow-RHS, a advanced link-prediction architecture. It splits into two parts: a left-hand side (LHS) device tower and a right-hand side (RHS) content tower. The LHS taps into temporally valid watch-history data, effectively capturing collaborative signals. Meanwhile, the RHS tower stays intentionally shallow, focusing solely on intrinsic content features. It sidesteps ID-based embeddings or neighbor aggregation, forcing the system to operate on available intrinsic content data. This approach ensures that both new and existing content are embedded into a collaborative-filtering-aware space.

Impact on Content and Device Engagement

The results are telling. Large-scale online experiments at Tubi show increased engagement and faster promotion of new content. The approach isn’t just limited to content, though. Tubi applies similar principles to solve device cold-start scenarios by constructing cohort-based embeddings from demographic data. Why does this matter? Because user engagement is the lifeline of streaming services. Slow engagement means lost opportunities, and Tubi’s method speeds up the entire process.

Here's what the benchmarks actually show: Consistent improvements in content cold-start engagement and device engagement. But is this the silver bullet for all streaming platforms? Not quite. It’s a significant step forward, but adaptation varies across different ecosystems. The architecture matters more than the parameter count here, making it a more adaptable, scalable solution.