How Recommendation Systems Work
A middle-school guide to recommendation systems: how YouTube, Netflix and Spotify suggest content using collaborative filtering, content-based filtering and feedback loops.
Key takeaways
- Recommendation systems predict what you might like by studying patterns in your behaviour and other people's
- Collaborative filtering uses people with similar tastes; content-based filtering uses the features of items
- Every click, watch time and skip is data that feeds the system
- Recommenders can trap you in a filter bubble, so it helps to explore on purpose
Why your screen knows you
Open YouTube, Netflix or Spotify and the home screen is already full of things picked just for you. No human chose them. A recommendation system did, in a fraction of a second. It is one of the most widely used kinds of AI in the world.
A recommender is a system that predicts what you are likely to enjoy and puts those items in front of you. The goal is usually to keep you watching, listening or shopping. To do that well, it needs data and a clever way to use it. If you are new to how machines learn from data, start with What Is Machine Learning?.
You are made of data
Every action you take is a tiny piece of information:
- Which videos you click
- How long you watch before leaving
- What you search for
- What you like, save or skip
- The time of day and the device you use
None of these signals means much alone. But across billions of actions from millions of people, clear patterns appear. The system turns your behaviour into numbers and looks for matches.
Two main strategies
Collaborative filtering
This is the "people like you" method. The system finds users whose behaviour is similar to yours, then recommends things they enjoyed that you have not seen yet.
Imagine a simple table of users and the films they rated. If you and another viewer both loved the same five films, and that viewer also loved a sixth film you have never watched, the system bets you will like it too. It never needs to understand why the films are good. It only needs the pattern of who likes what.
Content-based filtering
This method studies the items instead of the crowd. Each song, video or product is described by features: genre, tempo, length, topic, actors, keywords. If you keep choosing fast electronic music, the system recommends other tracks with similar features.
Most real services, like Netflix and Spotify, blend both methods and add many more signals. This is called a hybrid system.
The feedback loop
Recommenders learn continuously. You watch something, the system records your reaction, and it updates what it shows next. This is a feedback loop.
The loop is powerful but tricky. Because the system keeps showing what you already like, you can end up in a filter bubble, seeing the same kinds of ideas over and over and missing the rest of the world. A recommender can also chase watch time so hard that it favours dramatic or extreme content, because that keeps people clicking.
Using recommenders wisely
You are not helpless inside the algorithm. You can:
- Explore on purpose by searching for new topics and creators
- Use dislike and "not interested" buttons, which are strong signals
- Clear your history to reset what the system thinks
- Notice when you are stuck in a loop and step out of it
Understanding the machine puts you back in charge of it. For more on staying in control of AI tools, read Using AI Safely and Responsibly.
Quick quiz
Test yourself and earn XP
What does a recommendation system try to do?
A recommender predicts which items you are most likely to like, based on data about you and other users.
What is collaborative filtering?
Collaborative filtering finds users who behave like you and suggests what they liked but you have not seen yet.
What does content-based filtering use?
Content-based filtering looks at item features, like a song's genre or tempo, and recommends similar items.
Why might skipping a video matter to the system?
Skips, watch time and likes are all signals. A quick skip tells the system you were not interested.
What is a 'filter bubble'?
A filter bubble happens when a recommender keeps showing similar things, so you stop seeing new or different ideas.
FAQ
Reputable systems use signals like what you watch, click, search and how long you stay. They generally do not need to read private messages, though you should always check an app's privacy settings.
Yes. Liking, disliking, clearing watch history, and deliberately exploring new topics all change the signals the system learns from.
Keep exploring
More in AI