Recommendation System

A recommendation system is a computer program that calculates suggestions from huge amounts of usage data. It observes what you have clicked on, bought, or watched to the end so far. From this, it estimates what might interest you next. The basic idea is simple: people with similar behavior often like similar things. Whoever has watched the same five series as you has probably got a useful tip for you on the sixth as well. The system then sorts out the few options it shows you from millions of possibilities.

Why order determines success

Large platforms have a volume problem. On YouTube, several hundred hours of video are uploaded every minute. No human could search through this supply. Without sorting, such a service would be practically unusable. The recommendation system takes on exactly this pre-selection.

For providers, this involves a great deal of money. Netflix has publicly stated that a large share of the content watched is found through recommendations rather than search. At Amazon, a significant portion of revenue depends on the suggestion bars. Whoever controls the order, therefore, controls what gets sold and seen.

That is why such systems are also politically contentious. If a feed mainly shows what holds attention the longest, it may favor sensational or one-sided content. The EU’s Digital Services Act therefore obliges large platforms to disclose the main criteria behind their recommendations. Users must also be able to choose a view that is not tailored to their personal profile.

Of neighbors, features, and feedback loops

The classic method is called collaborative filtering. The system looks for users whose past ratings resemble yours. Whatever this group liked and you do not yet know is suggested to you. Nobody needs to describe their own preferences; they emerge from the comparison. The downside: for a completely new user, comparison data is missing. This starting problem is called the cold start.

The second family works with properties of the content itself, such as genre, language, length, or author. If you liked a science fiction audiobook, the system suggests more from the same category. Modern services combine both approaches and supplement them with neural networks, i.e., learning-capable computational models modeled on nerve cells. These models convert users and content into long strings of numbers. If two such number strings lie close together, they are considered similar.

The selection process almost always runs in two stages. First, a fast procedure filters a few hundred candidates out of the entire offering. Then a more precise model evaluates only this selection and puts it into an order. Afterward, the system measures whether you actually clicked. This feedback flows into the next round of training. This is exactly how a well-known error arises: what is placed at the top gets clicked on more often and thereby appears more popular than it actually is. Experts speak of a feedback loop.

From the Netflix home screen to credit approval

You encounter recommendation systems constantly in everyday life, usually without any indication. Netflix’s homepage is compiled differently for each household, sometimes down to the preview images. Spotify builds weekly playlists from this. TikTok is considered the example where almost the entire feed consists of such predictions, without you having to follow anyone.

In business news, these systems often appear indirectly. When a platform reports rising watch time or better conversion, an improved recommendation model is almost always behind it. Online retailers, news portals, and job boards also use them. Even banks use similar technology to offer suitable products.

A common misconception is that a recommendation system knows your interests. It only knows patterns in data and calculates probabilities. It does not know why you watched a video to the end. It only knows that you did so, and that similar users behaved similarly.