Data & AI
Recommendation systems
Your feed seems to know you better than your friends do. Not magic, just statistics. Here is how it works, and how to steer it.
You open an app and the third video is exactly your taste. Coincidence? No. Behind every feed, playlist and shopping suggestion works a recommendation system that predicts, from billions of interactions, what might hook you next. It does not read minds. It computes with your behaviour.
Signals instead of mind reading
Every action in an app is a signal: what you click, how long you watch a video, what you skip, whom you follow, even how long you linger on a post without liking it. From these signals a profile of your interests emerges. Signals that are hard to fake are especially valuable, above all watch time. A like can be polite, but eight minutes of watching rarely lie.
People like you
The second building block is called collaborative filtering: the system finds users whose behaviour resembles yours and recommends what those people enjoyed and you have not seen yet. If you and a thousand others like the same ten shows, and nine hundred of them also watch show eleven, show eleven gets suggested to you. The system does not need to understand you or the show. Similarity in the data is enough. The final selection emerges from many such decision steps: fits the profile or not, already seen or new, likely to keep you in the app or not.
A decision tree is nothing magical: just a chain of simple questions with an answer at the end.
The filter bubble
A system that mostly shows you what you liked before has a side effect: it shows you more and more of the same. Your interests appear bigger and more widespread than they are, and opposing views show up less often. This is called a filter bubble. It is not an evil plan but the logical consequence of a goal many platforms optimise for: your time spent. Whatever grips you gets amplified, and outrage unfortunately grips especially well.
Steering on purpose
The good news: the feed learns from you, so you can train it. Use 'not interested' and 'do not show this again' consistently, they work better than just scrolling past. Deliberately follow sources outside your bubble and search for important topics actively instead of waiting for them to wash up. Many platforms now offer a chronological feed or a history you can clear, and in the EU large platforms are even required to offer an option without personalisation. And the most effective reminder remains: your feed is a selection optimised for time spent, not a picture of the world.
Exercises
0 of 6 solvedTime to try it yourself. You can't break anything, every attempt counts.
What does a recommendation system mainly compute its suggestions from?
What does collaborative filtering mean?
You watch 3 out of 4 suggested cat videos to the end. What percentage of the suggestions is that?
A signal that is hard to fake and therefore especially valuable is ….
A decision path asks 3 yes or no questions in a row. How many different answer combinations are there?
Match each term to the description that fits it.