liveBook · Manning

Popularity bias

As explained in AI Recommender Systems

A recommendation bias in which popular items receive disproportionate exposure or recommendation frequency.

Popularity bias occurs when a recommender gives popular items disproportionate exposure or recommendation frequency. In a feedback loop, exposure generates interactions, interactions reinforce popularity, and the catalog's long tail becomes less visible. section 1.6 section 4.6.4

Popularity bias appears in several parts of a recommender system:

  • Collaborative filtering can repeatedly surface items that already have abundant interaction data. section 1.6
  • Low item coverage reveals that recommendations concentrate on a small portion of the catalog. section 4.6.4
  • Exploration competes with exploitation: proven popular items are safer choices, while less-known items receive fewer opportunities to generate evidence. section 1.6
  • Popularity bias can reduce aggregate diversity even when individual recommendations achieve strong precision. section 4.6.4

A coverage warning

An ALS model recommends only 2.6% of the available catalog, with more than half of those recommendations coming from the most popular quartile. Compare this with a popularity baseline, which has 100% coverage by recommending the same top-tier items to everyone. section 4.6.4

Figure 4.18 ALS recommends only a small percentage of available items.

Use the catalog-coverage figure to inspect how a personalized model concentrates recommendations on a small set of items. section 4.6.4

How AI Recommender Systems covers it

How deep it goes
Traces the mechanism from exposure to reinforced popularity and diagnoses it with coverage, while deferring mitigation strategies.
Examples
Uses a 2.6% catalog-coverage result, popular-quartile concentration, and a 100%-coverage popularity baseline.
What it assumes
Builds on interaction data, collaborative filtering, recommendation feedback loops, and the candidate-generation pipeline.
What you can do after
Diagnose popularity bias by comparing item coverage and concentration against a popularity baseline.
How it connects
Connects popularity bias to collaborative filtering, feedback loops, cold-start problems, item coverage, exploration, exploitation, and diversity.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Recommender Systems. The section links point to the parts of the book it draws on.