Recommendation narrowing

Why does a system built to show you what you like end up narrowing what you see?

A recommender has an unambiguous job: predict what you would enjoy, and put it in front of you. Nothing in that job description says narrow the world. And yet the complaint is universal — the feed converges, the suggestions start rhyming with each other, and the thing you loved last year keeps coming back dressed slightly differently. Where does the narrowing come from, if no one asked for it?

Start with the data the system learns from. It is not a survey of what you like; it is a log of what you clicked. And you could only click on what you were shown — which the system itself chose. So the training data is generated by the model being trained. Does that matter? Consider what it means for anything not shown: it accumulates no evidence at all. Not evidence of dislike — silence.

Now ask how a system should treat silence. It has to estimate an item's appeal from data, and for an item never surfaced the estimate is whatever the model's priors and its sparse similarity signals imply, held with enormous uncertainty. If ranking is done by expected score, the uncertain item loses to the item with a solid, measured, moderately-good record. It loses again tomorrow, for the same reason. The system never learns it was wrong, because being wrong would require showing it.

That is the trap in its cleanest form: the exploitation side of the exploration-exploitation dilemma, where an agent maximising immediate reward from current estimates stops gathering the information that would improve them. The remedies are all deliberate exploration — reserving slots for uncertain items, or ranking by an optimistic bound rather than a point estimate, so uncertainty itself earns a showing.

A second, blunter mechanism runs alongside. Popular items are shown more, so they collect more feedback, so their estimates are both high and precise, so they are shown more. Position compounds it — the top slot earns clicks partly because it is the top slot, and the log records that as merit. Neither of these requires anything about you; they would narrow a catalogue with no personalisation at all.

So does the loop close on the individual too? Partly. Your behaviour is shaped by what you were shown, and the next model reads that behaviour as your revealed preference. Whatever the recommender nudged becomes, on the record, who you are.

Now the honest part — because the strong version of this story has taken a beating. The "filter bubble", in the sense of algorithmic curation sealing people into ideologically narrow worlds, has not held up well when measured. Large field experiments run on Facebook and Instagram around the 2020 US election found that replacing the ranked feed with a chronological one changed what people saw substantially, but did not measurably reduce polarisation or change political attitudes over the study period. Other work repeatedly finds that people's own choices — who they follow, what they click, whom they talk to — narrow their diet more than the ranking does, and that recommenders sometimes widen an individual's range compared with what they would have found alone.

Hold both findings at once. The mechanism is real and demonstrable in the systems themselves. Its measured effect on people is smaller and messier than the popular account claims.

Three refinements tie the steps together.

First, the loop is about information, not preference. The system does not decide you are narrow; it simply stops sampling the alternatives, so its uncertainty about them never resolves. That is why exploration is a fix and better modelling alone is not.

Second, popularity bias and personal feedback are separable causes with different remedies. Debiasing by exposure — weighting observed feedback by the chance the item was shown — addresses one; forced exploration addresses the other.

Third, "narrowing" needs an object. Narrowing of an individual's diet, of the catalogue in active circulation, and of the range of views a society encounters are three different measurements, and evidence for one is not evidence for the others. Much of the disagreement here is people measuring different things and arguing about the answer.

One correction worth stating plainly: the narrowing is not a system working badly. It is what a system optimising short-horizon accuracy does when the data it learns from is the data it produced. Diversity, novelty and long-run retention have to be written into the objective on purpose, because accuracy will not imply them.

• How bandit algorithms price uncertainty so that exploring is not simply a cost.
• Why off-policy evaluation is hard: judging a ranker you never ran, from logs made by one you did.

Every term the collection defines is gathered in the glossary.