Instagram Reels recommendation system, explained simply

If an interviewer asks how Instagram picks the next Reel, "it looks at your likes" is the answer that loses points. A typical recommender works in stages, and each stage has one job.

Here's the short version. Build a shortlist, score it with several guesses, and keep learning from logs.

Start with a shortlist

The system can't score every Reel in the catalogue for every swipe. There are millions of them. So the first stage, candidate generation, narrows millions of Reels down to about a thousand that you might plausibly see.

Those thousand are the candidates. Think of it as a shortlist of Reels that could show up for you. Nothing is ranked yet. The only goal is to make the next step small enough to run.

One model, several guesses

Next comes the ranking model. It's a program that looks at each Reel on the shortlist and guesses four things. Will you watch it, will you like it, will you send it, and will you skip it.

That's the part people miss. The model doesn't predict a single thing, and it isn't only about likes. It makes several guesses per Reel.

Those guesses are combined into one score. The Reels with the top scores play first on your phone. Your feed order is just that list, sorted.

What happens when you skip

Your phone doesn't only receive Reels. It also reports back. What you watched, skipped and sent goes into watch logs, which work a bit like a ledger of your behaviour.

In a typical setup, the model keeps learning from those logs. So a skip isn't wasted. It becomes data that feeds the next round of guesses.

This also explains why the model guesses skip at all. Skipping is a signal, same as a like or a send, and it ends up in the logs like the rest.

The answer to give in an interview

Keep it to short beats. Generate a shortlist of candidates. Score each one using several guesses from a ranking model. Log what people do and keep learning from it.

The flow runs candidate generation, then the ranking model, then your phone, then watch logs. The logs loop back into learning, which is why the system keeps adjusting.

Next time you're sketching a recommender on a whiteboard, draw those four boxes in order and say what each one does. If you only remember one thing, remember that the model makes several guesses and doesn't rely on likes alone.