How Do Recommendation Algorithms Work? People Like You Liked This

The oldest and still most important technique makes no use of what an item actually is. It builds a very large and very sparse matrix of users against items, filled in where someone has watched, bought or rated something, and then predicts the empty cells. The intuition has two forms: find users whose history resembles yours and recommend what they liked, or find items that tend to be liked by the same people and recommend items similar to those you liked. The practical method for doing this at scale is matrix factorisation, which decomposes the huge matrix into two much smaller ones representing each user and each item as a list of perhaps a hundred numbers in a shared space of latent factors. Those factors are discovered rather than specified, and they often turn out to correspond loosely to recognisable qualities, so a dimension might separate gentle from intense or mainstream from obscure, although nothing guarantees they will be interpretable. The approach was popularised by the Netflix Prize, a competition running from 2006 to 2009 offering a million dollars for a ten percent improvement in rating prediction, which drew enormous academic attention and whose winning ensemble was reportedly too complex to deploy.

Modern systems combine several approaches, because each has a failure mode the others cover:

• •Content-based filtering uses attributes of the item itself, its genre, cast, tags, text or audio features, which lets a brand new item be recommended before anyone has interacted with it
• •Hybrid systems blend collaborative and content signals, which is what almost every large deployment now does
• •Deep learning models learn representations of users and items from raw interaction sequences, and treat recommendation as predicting the next item in a sequence, which is closely related to how language models work
• •Two-stage architecture is standard at scale: a fast candidate generator narrows millions of items to a few hundred, then a heavier model ranks those precisely, because scoring everything for everyone is computationally impossible
• •Contextual signals including time of day, device, location and what the user did in the last few minutes, which often matter more than long-term preference
• •Explicit ratings are largely obsolete, since implicit signals such as watch time, pauses, rewatches and how fast someone scrolls past are more abundant and more honest than stars

Several difficulties are structural rather than bugs. The cold start problem affects any new user or new item with no interaction history, and is why services ask you to pick a few interests when you sign up. Popularity bias is pervasive, since popular items accumulate more interactions, which makes them more recommended, which makes them more popular, and correcting for it requires deliberate intervention. The filter bubble concern, that a system optimising for engagement narrows what a person sees, is real in mechanism and more contested in measurement, with studies finding effects that are smaller and more variable than the popular account suggests. Feedback loops are the deepest issue: the system's recommendations determine what data it collects next, so it trains on the consequences of its own decisions and can never observe what would have happened if it had recommended something else, which makes evaluation genuinely difficult and is why serious operators run continuous randomised experiments.

The most consequential question about any of these systems is what objective it has been given, because it will pursue that objective faithfully and nothing else. A system optimised for click-through will learn that provocative and misleading material gets clicked. One optimised for time spent will learn that outrage and autoplay hold attention. One optimised for immediate engagement will under-recommend things a person would have valued more and clicked less, which is why several large platforms have shifted towards predicting longer-term satisfaction, explicit user surveys and measures such as whether someone returns the following week. The gap between the metric and the actual goal is the recurring failure, and it is a specification problem rather than a technical one: the system is not misbehaving when it maximises watch time with material nobody would endorse on reflection, it is doing exactly what was asked.

Regulation has begun to reach these systems, with European rules requiring very large platforms to explain the main parameters of their recommenders, to offer at least one option not based on profiling, and to assess systemic risks. The practical levers available to a user are narrower and real: watch history can usually be edited, and since the system infers preference from behaviour, a single curious click can shift recommendations for weeks; most services allow a non-personalised or chronological view somewhere in the settings; and separate profiles prevent one household's viewing from contaminating another's. The wider point is that these systems are extremely good at finding more of what you already engage with and structurally poor at finding what you would value but have never encountered, because nothing in the training data tells them about that. Deliberate exploration, through a curated source, a recommendation from a person, or simply a list of things nobody is being paid to show you, remains the only reliable way to get outside the loop.

Recommendation systems work mainly by collaborative filtering, predicting what you will like from patterns across many users without understanding the items at all, usually by factorising a huge sparse matrix into compact representations of users and items. Modern deployments blend this with item attributes, sequence models and context, and use a fast candidate stage followed by precise ranking. Cold start, popularity bias and self-reinforcing feedback loops are structural, and the system's behaviour follows entirely from the objective it was given.