Recommender System

A recommender system is a program that makes suggestions. It observes what you have watched, bought, or clicked away from so far. From this, it estimates what might interest you next. When a streaming service offers you a series on the homepage, such a system has made that selection. The German term for this is Empfehlungssystem, meaning the same thing. Important: the system doesn’t know you personally, it only calculates based on your behavior and that of millions of others.

Whoever fills the homepage steers attention

The range of content on the internet is larger than any human could ever browse through. A large online shop carries hundreds of millions of items. A video platform receives several hundred hours of new material per minute. Without pre-selection, this would be unusable. Recommender systems are the filter that turns this mass into a few dozen suggestions.

For companies, a great deal of money depends on this. For streaming services, the majority of titles actually watched come from recommendations rather than from search. For online shops, good suggestions noticeably increase revenue per visit. That’s why companies like Amazon, Spotify, or TikTok invest heavily in this technology. Many even see TikTok as an example of how a particularly good recommender system alone can make a product successful.

There is also a societal side to this. If a system mainly shows content that holds attention for a long time, it may favor the exciting and the extreme. Critics speak of filter bubbles: you see more and more of the same. That’s why the EU, in its Digital Services Act, requires large platforms to disclose the criteria by which they recommend content. Users must also be able to choose a version there that is not tailored to their profile.

Of neighbors, features, and click probabilities

The classic approach is called collaborative filtering. The idea behind it is simple: people with similar tastes like similar things. So the system searches for users whose past ratings match yours. What these neighbors liked and you don’t know yet gets suggested to you. For this, no one needs to know what a movie is actually about in terms of content.

The second approach is called content-based filtering. Here, each item is described by features, such as genre, language, release year, or price range. If you liked three Scandinavian crime series, you get a fourth. In practice, combining both yields the best results. Modern systems additionally use neural networks, i.e. learning-capable computational models that extract suitable features from raw data themselves.

Technically, this usually runs in two stages. First, a fast step selects a few hundred candidates from millions of items. Then a more precise model ranks these candidates by estimated click or purchase probability. A well-known problem is called cold start: there is not yet any data on new users and new products. In that case, one relies on general trends or a few details from the sign-up process.

From the supermarket shelf to the news app

In everyday life, you encounter this technology daily, often without any indication of it. The order of your social media feed, the suggestion list after a video, the “customers also bought” section in the shop: all of these are recommender systems. Music services also use it to compile weekly playlists. Some news portals sort their homepage in the same way.

In business news, the term usually appears in three contexts: quarterly earnings reports from platforms, regulatory debates, and privacy issues. A common misconception is that these systems would need to listen in on conversations in order to guess so accurately. In reality, clicks, dwell time, and comparisons with millions of similar profiles are enough. One should also not confuse recommender systems with search engines: with search, you formulate a wish yourself, whereas with a recommendation, the system estimates it without being asked.