How would you approach the problem of recommending new or less popular items?
Instruction: Describe methods to mitigate the bias towards popular items in recommendation systems.
Updated
Example Answer
I would make eligible new or less popular items available to the candidate stage using metadata and content features, then test bounded exploration or reranking. Less popular does not automatically mean useful, so relevance and user controls still matter. Define the popularity measure explicitly: if rank one means most popular, larger ranks indicate less popular items. I would track qualified exposure, new-item engagement and coverage while checking that relevance and satisfaction do not regress.
Make it your own
Illustrative metric definition: popularity is prior-window interaction count; report exposure by popularity bucket and item age. Keep new items separate from older items with persistently low interest.
Why this works
The measure avoids a reversed rank interpretation and separates lack of evidence from poor relevance. Follow-up: how would you prevent exposure feedback from permanently burying a new item?
Sources: Google recommendation re-ranking, Google collaborative filtering.