Echo Chamber

A recommender learns from behaviour it produced itself. Show a user a narrow slice of the catalogue, learn from what they click inside that slice, and the next slice is narrower still. An echo chamber is where that loop ends up: a population whose options have been trimmed to what it already agreed with.

The loop that creates it
This is a closed-loop training artefact rather than a modelling error:

• the model ranks from a user's history, so unvisited regions of the catalogue receive no impressions;
• no impressions means no new interactions for those items, which training reads as absence of interest;
• the next round fits a more confident version of the same preference, and popularity and selection bias amplify it further.

Echo chamber or filter bubble
The two get blurred. A filter bubble is the individual outcome — one person's narrowed feed. An echo chamber is the social one: like-minded users reinforcing each other through shared recommendation, so a community converges and contrary content is crowded out rather than merely unseen. They are measured differently too — topic drift within a user for the bubble, homogeneity between users for the chamber.

Why it matters
Offline accuracy rewards the loop, since predicting more of what a user already consumed is exactly what it scores highest. Every countermeasure costs short-term relevance: diversity-aware re-ranking, an exploration budget in candidate generation, popularity debiasing in training, and tracking coverage and topic entropy per user over weeks instead of per request.