Is popularity bias the same as position bias?

No. Position affects attention because of where an item appears; popularity bias concerns overrepresentation of already popular items. They can interact.

Does it affect only commercial products?

No. It can affect recommendations for information, media, courses and other items.

Can personalization still contain popularity bias?

Yes. A personalized system can overemphasize common items and under-serve particular interests.

AI Glossary / Governance, ethics and safety / Intermediate

Popularity Bias

Definition

Popularity bias is a tendency for already popular items to receive disproportionate exposure while less popular items receive little representation. In recommendation systems, it can narrow discovery and produce suggestions that poorly reflect users whose interests extend beyond the most common items.

What it means

Popularity can be useful information. A widely used product or frequently read article may genuinely suit many people. The concern is not that popular items appear at all, but that the recommendation process overemphasizes them relative to its intended purpose and the interests it should serve.

Historical interactions are partly shaped by what people had a chance to see. An item that was displayed prominently can accumulate more clicks, which may then be treated as evidence that it deserves still more exposure. Less visible items provide fewer observations, even when some users would value them.

The effect matters to both users and suppliers. Users may receive repetitive suggestions, while smaller creators or specialist products struggle to be discovered. However, equal exposure for every item is not automatically the right objective. Relevance, quality, availability and the service's purpose still matter.

How it works

A recommender learns from signals such as views, purchases or ratings. Popular items often have more observations and may be favored by simple popularity rankings or by learning methods that fit frequent interactions well. Their continued exposure can generate another round of interactions, creating a feedback loop.

Assessment therefore considers exposure as well as interaction counts. Teams can examine how recommendations are distributed across items, how much of the catalog is reached and whether users with different preference patterns receive appropriate suggestions. A single aggregate accuracy score may conceal uneven experiences.

Possible interventions include adjusting training weights, collecting better evidence about less exposed items or re-ranking candidate recommendations. Each changes a different part of the process. Evaluate whether the intervention improves useful discovery without making recommendations irrelevant. A popularity measure, its time window and the comparison baseline should be explicit; different definitions can lead to different conclusions.

Three practical examples

A book recommender repeatedly suggests widely purchased titles to a reader whose history shows a strong interest in specialist publications.
A marketplace ranks products heavily by past interaction count, giving new or niche items few opportunities to accumulate evidence.
An internal learning portal displays the most frequently opened courses, while suitable specialist training remains difficult to discover.

Think of it this way

Think of a shop placing yesterday's best sellers beside the entrance every morning. They receive more attention partly because they are easy to see, and the resulting sales support placing them there again.

The analogy does not prove those products are unsuitable or that shelf space should be equal. It illustrates why observed popularity and opportunity for exposure can reinforce each other.

Worked example

Comparing counts with opportunities

In an invented recommendation test, Item A receives 1,000 displays and 100 clicks. Item B receives 100 displays and 15 clicks. Ranking solely by clicks favors A, with 100 compared with 15.

The observed click-through rates are 10% for A and 15% for B. These rates suggest that raw counts omit an important difference in exposure. They still do not establish that B is universally better: the items may have reached different users, appeared in different positions or been shown in different contexts, and B has fewer observations.

A team can investigate with a suitable evaluation that controls or accounts for those factors. The lesson is to distinguish popularity counts from evidence of relevance, not to replace one unqualified ranking rule with another.

Illustrative case study

Fictional case: Alder Courses notices that its learning assistant recommends the same introductory modules to experienced users. The training data contains many interactions with those modules because they were previously required during onboarding.

The team examines recommendations by users' learning goals and prior experience. It tests a re-ranking approach that can include relevant specialist material, while monitoring whether the suggestions still help people find suitable courses. It also records the special exposure created by the old onboarding policy.

This imagined case demonstrates how product history can shape recommendation data. It makes no claim about a measured improvement or a universally effective mitigation.

Common misconceptions and mistakes

Recommending a popular item is always biased. Popularity can be appropriate evidence; the concern is unjustified overemphasis and its consequences.
Low interaction proves low interest. People may never have encountered the item.
Promoting every unpopular item fixes recommendation quality. Less popular does not automatically mean relevant, useful or suitable.

Business use

Businesses can compare exposure concentration with their intended discovery goals and user needs. Review whether the system serves people with specialist interests as well as those who prefer mainstream options. Where suppliers depend on visibility, consider their perspective alongside user relevance.

Choose mitigation based on the mechanism observed. Re-ranking cannot reveal a candidate that the retrieval stage never supplies, while collecting more clicks under unchanged exposure may repeat the same imbalance. Evaluate the complete recommendation pipeline and report the tradeoffs rather than promising that one diversity score proves fairness.

Check your understanding

A niche item receives fewer clicks than a mainstream item but was shown one-tenth as often. Can the team conclude that users prefer the mainstream item?

Answer: Not from raw click counts alone. It needs to examine exposure, audience, position and other factors affecting the comparison.

Related terms

Feedback loop
Re-ranking
Recommendation system
exposure bias
position bias
long-tail recommendation

Learn next

Read exposure bias and feedback loop. For one recommendation report, place displays alongside clicks and identify which conclusions require a more controlled comparison.

Feedback loop

Sources

Artificial Intelligence Ontology, Popularity Bias: https://raw.githubusercontent.com/berkeleybop/artificial-intelligence-ontology/main/aio.owl. The returned definition links to NIST SP 1270. Abdollahpouri and colleagues, The Unfairness of Popularity Bias in Recommendation: https://arxiv.org/abs/1907.13286. Abdollahpouri, Burke and Mobasher, Managing Popularity Bias in Recommender Systems with Personalized Re-ranking: https://arxiv.org/abs/1901.07555. Relevant primary abstracts were reviewed; examples are original.

Previous: Pooling Layer Next: Population