A new subscriber opens a streaming app for the first time. You know their country, their device, and maybe what they clicked on the sign-up page. You don't know what they like. And this is the moment that matters most, because if the first screen feels irrelevant, they might not come back.
That's the cold start problem. The recommendation system needs data to work, and new users haven't given it any. The same is true for new content: a title that just arrived has no viewing history, so the system doesn't know who to show it to.
On a travel streaming platform like Travelxp, both problems showed up constantly. New subscribers came from many countries, and new content arrived regularly. These are the approaches I'd consider.
For new users
| Approach | How it works | Trade-off |
|---|---|---|
| Ask at sign-up | Let people pick a few interests or titles | Adds friction, answers may be rough |
| Use context | Country, device, time of day, sign-up source | Weak signal, but available instantly |
| Popular in your region | Show what similar users watch most | Safe but generic |
| Editorial picks | Curated rows chosen by people | Good quality, doesn't personalise |
| Learn fast | Weight early behaviour heavily | Can overreact to one accidental click |
In practice, most services combine several. A sensible default home screen built from editorial and regional popularity, adjusted quickly as the person starts watching.
Keep the sign-up questions light
Asking new users what they like can work well, but only if it's quick. Three to five choices from a visual grid, with a clear option to skip. Don't make it a long survey.
And use the answers visibly. If someone picks "food" and "mountains" and the home screen shows neither, they'll wonder why you asked.
Context is underrated
Even with no viewing history, context tells you something. Someone on a TV in the evening probably wants something different from someone on a phone at lunchtime. Someone in a particular country may be more interested in nearby destinations or local-language content.
These signals are rough, but they're better than showing everyone the same screen.
For new content
New titles have the opposite problem. You know a lot about the content but nothing about who'll watch it.
- Use the content's attributes: genre, destination, format, length. Show it to people who watched similar things.
- Give new content a fair chance by reserving some space for it, rather than letting popular older titles dominate.
- Watch early performance closely and adjust.
Google's recommendation systems course explains the difference between content-based and collaborative approaches, which is basically the difference between these two cold starts. 1
Measure the first sessions specifically
Overall recommendation metrics are dominated by existing users with lots of history. To see whether cold start is working, look at new users separately:
- How many start a play in their first session?
- How long does it take them to find something?
- How many come back within a few days?
These are among the strongest early signals of retention I know. The churn analysis post covers what happens if they don't come back, and the streaming metrics post covers the basics of measuring plays.
Don't overfit to the first click
If someone watches one cooking show by accident, the next home screen shouldn't be all cooking shows. Early signals should count, but not so much that one click takes over. Mixing in a bit of variety for new users helps the system learn what they actually like.
Simple first, smarter later
You don't need a sophisticated model to handle cold start reasonably. A good editorial default, sensible use of context and a quick reaction to the first few plays will get you a long way.
Build the more advanced approaches once you have enough users to learn from. A clever model trained on too little data often does worse than a thoughtful default.