Popularity Bias
A recommender trained on what people clicked will recommend what people clicked. Clicks concentrate on a small head of the catalogue — and items only collect clicks if they were shown — so the system inherits that concentration and then amplifies it. Popularity bias is that drift: already-popular items being over-recommended relative to how well they actually match a user.
Why it compounds rather than settles
The loop is closed. A popular item is shown more, collects more interactions, looks stronger in the next training run, and is shown more again. Nothing in an accuracy objective resists this, because the popular item usually is a decent bet — the error is in the margin, not the direction. Implicit feedback makes it worse: an unobserved item may be unseen rather than unwanted, and popularity-weighted negative sampling bakes the head in further.
How it shows up
- Flattering offline metrics — recommending the top 100 items to everyone scores respectably on Precision@k and terribly on coverage, so accuracy alone hides it.
- The long tail starves — niche items never accumulate the interactions they would need to become recommendable.
- Personalisation collapses — different users receive near-identical lists, and the product stops being a recommender.
What is done about it
Measurement first: track coverage, novelty and the popularity distribution of what you actually serve, next to the accuracy numbers. Then correction — propensity-based reweighting of training data, popularity-debiased sampling, or a re-ranking pass that caps head items per slate. Each trades a little measured accuracy for catalogue reach, which is a product decision rather than a modelling one.
Popularity bias is a recommender's tendency to over-recommend already-popular items beyond what a user's preferences justify, because training data records interactions only with items that were shown, and showing drives further interaction. The loop is self-reinforcing and invisible to accuracy metrics — serving the catalogue's head scores well on Precision@k while coverage collapses. It starves the long tail and erodes the personalisation the system exists to provide.
A bookshop puts the bestsellers by the door. They sell more because they are by the door, so next month they earn that spot again — and the quieter books never get a chance. Popularity bias is the same thing happening inside recommendation software. It learns from what people clicked, and people mostly clicked what they were shown, so whatever is already popular keeps getting pushed. The result feels oddly samey: everyone is handed roughly the same suggestions, and the unusual film or the small artist you might have loved stays invisible.