Solving Cold-Start with Multimodal Embeddings: Netflix's MAPS System
Explain how multimodal embeddings (CLIP, MediaFM) solve the recommendation cold-start problem, and why content-aware embeddings and model consolidation had to be combined — not tested separately — to produce a real production win.
4 sections · 9 lessons
Course outline
The Cold-Start Problem in Recommendation Systems
- Why ID-based models fail on new items
- What CLIP embeddings are and why they transfer taste signals
From Content-Awareness to Model Consolidation
- How CLIP embeddings enable merging per-canvas models into one
- Reward-based weighting to balance uneven data sources
Why Two Fixes Failed Alone but Won Together
- Ablation testing and the second blocking factor trap
- Real-world stress test: absorbing a sudden distribution shift
Reusing Embeddings Across Search, Video, and Infrastructure
- Query-aware ranking reusing the same embeddings for free
- MediaFM and multimodal video embeddings
- The Embedding Store: decoupling foundation models from downstream models