Free sample lesson · Solving Cold-Start with Multimodal Embeddings: Netflix's MAPS System
Why ID-based models fail on new items
Lesson 1 of the Wondering course Solving Cold-Start with Multimodal Embeddings: Netflix's MAPS System
Titles as Opaque IDs
Netflix's older asset personalization models did something surprisingly blunt.
Each artwork or video preview was treated as an opaque ID. Just a number in a lookup table.
The model had no notion of content. It did not know the colors, the faces, or the mood of the image.
It only knew how people had behaved toward that ID before.
- Stored
- Ignored
Learning Only From History
If an ID carries no content, the model can only learn from interactions.
Every click, play, and skip attached to that ID becomes the model's entire understanding of it.
This works well for a title that has been on the service for months.
The history is deep, so the pattern is clear.
- Assign ID: Each artwork or video preview is stored as an opaque ID.
- Collect clicks: Members interact with that asset and the system records it.
- Accumulate history: Interactions build up into a usable pattern for that ID.
- Personalize: The model uses that history to pick assets for members.
The New Title Problem
Now add a brand new title to the catalog.
New titles had no interaction history. Nobody had clicked, played, or skipped the artwork yet.
The model looks up the ID and finds nothing.
With no content understanding to fall back on, it has no basis at all for a personalized choice.
Falling Back On Popularity
So what did the system actually do?
It fell back to popularity heuristics. Show the broadly appealing option and hope for the best.
That is a reasonable safety net. It is also not personalization.
Every member effectively sees the same choice until the data arrives.
- New asset: An artwork or video preview enters the system as an opaque ID.
- History available?: The system checks for past interactions on this ID.
- Popularity heuristics: Everyone sees the broadly appealing option.
- Personalized pick: The model uses accumulated interactions for that ID.
- data accumulates
The Classic Cold-Start Problem
This has a name: the cold-start problem.
A system cannot serve an item well until enough data accumulates about it. The item starts cold and slowly warms up.
The waiting period is exactly when a new title needs strong presentation most.
So the weakness lands at the worst possible moment.
- Established title
- New title
Why This Design Fails
The root cause is not a lack of clever math.
It is the representation. An opaque ID carries zero transferable information.
Two pieces of artwork can look nearly identical, yet the model cannot connect them. Each ID stands alone.
Nothing learned about one asset helps predict another.
- Opaque ID representation: Each asset is stored with no notion of content.
- No transferable information: Each ID stands alone in the model.
- Similar artwork unlinked: Two near-identical images cannot be connected.
- History is the only signal: The model learns only from past interactions.
- Cold start on new titles: New titles have no interaction history to learn from.
- Popularity fallback: The system serves broad heuristics until data accumulates.
Key Takeaways
Old asset models saw IDs, not content.
New titles had no history, so the system fell back to popularity until enough data accumulated.
That is the classic cold-start problem in one sentence.
The fix points in an obvious direction: give the model a way to understand the asset itself.
Check yourself
Does an opaque ID tell a personalization model anything about an artwork’s colors, faces, or mood?
- Yes, it describes the content
- No, it only identifies the item
Answer: No, it only identifies the item
No. An opaque ID is only a lookup-table identifier; it carries no information about the asset’s content.
Older asset models represented each title as an _____. They learned from _____; when that history was absent, they used _____, creating the _____.
Answer: Older asset models represented each title as an opaque ID. They learned from interaction history; when that history was absent, they used popularity heuristics, creating the cold-start problem.
An opaque ID contains no content information, so the model depends on interaction history. For a new title, popularity heuristics fill the gap, producing the cold-start problem.
A new title is added to a catalog, and its artwork has received no clicks, plays, or skips. What is the most likely behavior of an ID-based personalization model?
- It selects a broadly popular option
- It transfers learning from visually similar artwork
- It analyzes the artwork’s colors and faces
- It infers the member’s preference from the new ID
Answer: It selects a broadly popular option
Because the new ID has no interaction history and the model cannot interpret the artwork, it falls back to a broadly appealing popularity-based option rather than making a personalized choice.
Keep going
This is lesson 1 of 9 in Solving Cold-Start with Multimodal Embeddings: Netflix's MAPS System. Next up:
- What CLIP embeddings are and why they transfer taste signals
- How CLIP embeddings enable merging per-canvas models into one
- Reward-based weighting to balance uneven data sources
- Ablation testing and the second blocking factor trap
- Real-world stress test: absorbing a sudden distribution shift
Wondering turns any topic into short lessons you can read, watch, or listen to, with practice questions and an AI tutor.