How Recommendation Systems Decide What You See
1. Quick Summary
The classic approach is collaborative filtering: build a huge table of who interacted with what, find users whose behaviour resembles yours, and recommend the things they engaged with that you have not seen. The content of the item barely enters into it.
Modern systems add content features and learned models on top, but the core logic survives: your future behaviour is predicted from the behaviour of similar people, which works well and has predictable failure modes.
2. What It Means
The original formulation is simple. Represent every user as a vector of interactions, measure similarity between users, and for each unwatched item compute a weighted average of how similar users rated it. That is a neighbourhood method, and it is still a reasonable baseline.
Matrix factorisation improved on it. Instead of comparing users directly, the system learns a small set of latent factors for each user and each item, so that the predicted rating is the match between them. The factors are not labelled by hand; they emerge from the pattern of behaviour and often end up resembling genre, mood or pace.
Content-based methods exist for the case where behaviour data is missing. A new item with no interaction history can still be recommended if its features match items a user already likes, which is why most production systems combine both approaches.
3. Why It Happens
The cold start problem shapes much of the design. A new user has no history, so the system has to ask, infer from context, or fall back on popularity. A new item has no audience, so it needs content features or deliberate exposure before the behavioural signal can take over.
Feedback loops are the structural problem. The system recommends, you engage, the system learns from that engagement and recommends more of the same, which narrows what you are shown over time. This is not a bug in any particular algorithm, it is what optimising for engagement does.
There is also a permanent tension between exploitation and exploration. Showing what is most likely to satisfy uses the signal you have; showing something unexpected gathers the signal you need. Too much of the first produces stagnation, too much of the second produces irrelevant results.
4. Real Examples
Video platforms are the clearest case. Watch a few items on one topic and the recommendations narrow rapidly within a session, sometimes within a handful of clicks, because recent behaviour is weighted much more heavily than historical preference.
Shopping recommenders show the complementary pattern. People who bought this also bought that works well for items that are functionally related, and it fails in the familiar way when a purchase was a gift rather than a preference.
Music and reading services face a different problem because repetition matters. Recommending an item you have already consumed is sometimes exactly right and sometimes annoying, which is why those systems model whether an item is worth repeating rather than only whether it is worth trying.
5. How It Affects Us
The objective function is the real decision. A system optimised for time spent will behave very differently from one optimised for whether you return tomorrow or whether you say you were satisfied, and the differences show up in what gets promoted.
The narrowing effect has been documented repeatedly, and it is strongest for content that is engaging but not something the user would have chosen deliberately. Because engagement is easy to measure and satisfaction is not, the measured objective tends to win.
For users, the practical defence is awareness and deliberate input. Ratings, follows, explicit dislikes and clearing history all change the input data, and since the model is inferred from behaviour, changing behaviour is genuinely the way to change what it does.
6. Key Takeaways
- Collaborative filtering predicts from similar users, not from item content, which is why it works without understanding the item.
- Matrix factorisation learns latent factors from behaviour patterns and generally beats direct comparison.
- Feedback loops narrow recommendations over time because the system learns from what it already showed you.
- The objective chosen, engagement versus satisfaction, determines behaviour more than the algorithm does.