Filter Bubble
A recommender that serves you well is also, quietly, narrowing what you see. The filter bubble is the state a user ends up in when personalisation has removed so much of the catalogue that their view of it is shaped mainly by their own past behaviour.
How the loop closes
The mechanism is a feedback loop, not an intention. A model ranks by predicted engagement, so it shows more of what you engaged with; you can only engage with what was shown; the next training round reads that as confirmation. Each turn sharpens the estimate of your existing taste and leaves the rest of the catalogue unobserved. Popularity bias and selection bias drive the same loop from the item side, so the narrowing is reinforced from both ends.
Why offline metrics hide it
A bubble is invisible to the usual measurements. Held-out NDCG and recall@k reward predicting the items a user actually interacted with — precisely the items the previous policy promoted — so a narrowing system can score better and better while the experience degrades. Catching it needs metrics about the list served rather than its accuracy:
- Coverage: the share of the catalogue that reaches anyone at all.
- Diversity within a single slate, and novelty relative to a user's own history.
- Drift in the topical spread of one user's recommendations over weeks.
What is done about it
The remedies work against the objective on purpose: re-ranking a slate for diversity, exploration that spends impressions on unproven items, and caps on how much of a page one topic may occupy. Each costs short-term engagement, which is why a filter bubble is a product decision as much as a modelling one.
A filter bubble is the narrowed view of a catalogue that results when personalisation keeps serving items close to a user's past behaviour, so alternatives are never shown and therefore never evaluated. It arises from a feedback loop in which the model learns from interactions it generated itself. Accuracy metrics cannot detect it — coverage, diversity and novelty can — and the standard remedies are diversity re-ranking and deliberate exploration.
Imagine a newsagent who notices you buying one magazine and, trying to be helpful, stops stocking the others. Soon the shelf holds nothing but variations of your old choice. Nothing has been refused to you — you simply never see the alternatives to ask for them.
That is a filter bubble. Apps that pick what to show you learn from what you tapped before, and what you tapped before could only ever come from what they had already shown you. The loop tightens with every round. The fix is for the system to deliberately slip in things outside your usual pattern, even though you are a little less likely to click them.