How Algorithms Curate: Recommendation Systems, Objectives and Feedback Loops

Remember from Grade 11

In Grade 11 you learned that algorithmic personalization means two people can see different versions of the media world, and that platforms reward engagement. In the first lesson of this unit you met the ranking system as a first audience and gatekeeper.

This lesson opens the box (as far as anyone outside the companies can). You will explain how recommendation systems work in general, how they shape audience responses (1.4), how they reflect industry goals (1.6), and how they act as part of a medium's form (2.1).

Hook: three days to a different feed

Try this thought experiment. A new account watches two cooking videos all the way through, skips everything else, and rewatches one clip of someone flipping a crêpe. Within a few days, the feed may be mostly food. Nobody at the company chose that for you. A system predicted it, from your behaviour and from the behaviour of millions of other people.

How recommendation systems work

Platforms such as video apps, music services and shopping sites use recommendation systems because there is far more content than anyone could see. In general terms:

1. Signals: the system records behaviour: what you watch and for how long, replays, likes, shares, comments, follows, searches, skips, and sometimes your device, location and time of day.
2. Candidates: from millions of items, it quickly picks thousands that might suit you, for example items liked by people with similar behaviour (collaborative filtering) or items similar to what you liked (content-based recommendation).
3. Prediction: a machine-learning model estimates the probability that you will do each thing (finish, share, comment) for each candidate.
4. Ranking: it combines those predictions into a score using weights chosen by the company, and sorts.
5. Feed: you see the top items, plus some exploration (new kinds of content to test your interest).
6. Feedback loop: your reactions become new signals.

Companies do not publish their full systems, so outside researchers study them through experiments, leaked documents and company statements. Treat confident claims about exactly how the algorithm works with caution.

Why objectives matter

The objective is what the system is optimized for. Many feeds have been optimized largely for engagement (time spent, interactions), because attention is what platforms sell to advertisers. Content that triggers strong emotion, including outrage, often generates comments and shares, so an engagement objective can amplify it, even if viewers later say they didn't enjoy it.

Companies can change the weights, add signals such as user surveys, or demote content they classify as low quality. Every one of these is a choice about values, made by an industry with commercial interests.

Effects on audiences

• Filter bubble: Eli Pariser's 2011 term for a personalized world where you mostly see what fits your past behaviour. Echo chamber describes communities where people mostly hear views like their own. Research findings are mixed: many people's media diets are more varied than the fear suggests, but personalization can still narrow what some users see, especially on niche topics.
• Creators adapt: they design hooks, lengths and posting times to suit the system. Some use algospeak (substitute words such as unalive) to avoid automatic moderation.
• Different responses: two viewers may react differently to a topic partly because their feeds have shown them different examples of it for months.

Regulation and choice

Some governments now regulate recommendation systems. The European Union's Digital Services Act requires the largest platforms to offer at least one feed option not based on profiling users. Many apps also let users reset recommendations, mark not interested, or choose a chronological following feed.