Topics Technology and the internet

How do recommendation algorithms work?

A recommendation algorithm is a program that guesses what you will want next and puts it in front of you. It does this by looking at what you have clicked, watched, bought or skipped, and at what millions of other people did, then scoring thousands of possible items and showing you the highest scorers. The details differ between a video app, a shop and a music service, but the core idea is the same: prediction from patterns.

What makes it interesting is how little the system needs to understand. It does not have to know what a song is about. If people who played the same tracks as you also loved a certain album, that is enough to suggest it. Two classic approaches, comparing people with similar tastes and comparing items with similar features, are usually blended, and modern systems add machine learning on top. Companies rarely publish their exact recipes, so outside descriptions are partly informed guesses.

An episode on this would walk through the pieces: the signals, the candidate list, the ranking, and the side effects such as filter bubbles, which are real but debated in how strong they are. bre's hosts are AI, so they can be wrong, and you can press Talk to ask them to slow down or explain a step.

What a bre episode would cover

An outline of the episode bre would make for this question. Every episode is written fresh when you ask, so yours will differ.

  1. The question every feed is answeringWhat a recommender is trying to predict, and why a ranked list beats a pile of everything.
  2. Signals: what you leave behindClicks, watch time, skips, purchases, searches and ratings, and why quiet behavior often counts more than stated taste.
  3. People like you: collaborative filteringHow matching your history with similar users lets a system recommend things without knowing what they are.
  4. Items like this: content-based methodsUsing the traits of things you liked, such as genre, tempo or topic, to find close relatives.
  5. Candidates, then rankingWhy systems first narrow millions of items to a few hundred and then score those more carefully.
  6. The cold start problemWhat a service does when you are new or an item has no history, and why first impressions are guesswork.
  7. Bubbles, goals and trade-offsWhat it means to optimize for engagement, what is debated about filter bubbles, and how to nudge your own feed.

How the episode might open

A sample exchange between two of bre’s AI hosts, bre and Arlo. Both are AI; this is written by AI, as every bre episode is.

  1. breAI host

    Picture opening a video app and the first thing on screen is oddly perfect. Nobody picked that for you by hand. So what actually happened in the half second before it appeared?

  2. ArloAI host

    A guess. Software guessed.

  3. breAI host

    A very educated guess. It looked at what you watched, how long you stayed, what you skipped, and what people with a similar trail did next.

  4. ArloAI host

    So it doesn't know me. It knows my pattern.

  5. breAI host

    Right, and that distinction matters. It doesn't need to understand the video at all. It only needs to notice that people who liked A, B and C kept going to D.

  6. ArloAI host

    Who counted that? How does it even know they liked it? Nobody clicks a heart on everything.

  7. breAI host

    Good question, and it's where it gets interesting. Mostly it doesn't ask. It watches what you do, like whether you finished or bailed in ten seconds.

  8. ArloAI host

    Fine. Behavior over opinions. Makes sense.

Questions people also ask

Do recommendation algorithms listen to my conversations?
There is no solid public evidence that major apps use your microphone to pick recommendations, and companies generally deny it. What they do have is a lot of ordinary data: searches, clicks, location, and what people near you or like you do. That often feels uncanny without any listening.
What is collaborative filtering?
Collaborative filtering recommends items based on the behavior of similar users. If people whose history resembles yours enjoyed a certain film, the system may suggest it to you. It works without knowing anything about the film itself, only who watched what.
Why do I keep seeing the same kind of content?
Systems tend to show more of what you already engaged with, because that is the safest prediction. How strongly this narrows what people see is debated by researchers. Searching for new topics, skipping, or using not-interested controls usually changes what the system learns about you.
Can I reset or change my recommendations?
Many services let you clear watch history, pause personalization or mark items as not interesting. These controls change the signals the system uses, so the feed typically shifts over time. Exactly what each setting does varies by service and is described in its help pages.

Related topics

More: all 300 topics, technology and the internet, or the longer reads on /learn.

bre’s hosts are AI, and every episode is generated, so they can be wrong: check anything that matters. This page outlines what an episode would cover. It is for interest and learning, not medical, financial or legal advice.