RecSys interview questions

Why did you choose a pointwise loss rather than a listwise loss for the ranker?

Риксис. Какие вообще есть метрики там на точность рекомендаций?

Which ranking metrics do you know?

How is MAP (Mean Average Precision at K) calculated?

What metrics other than accuracy-related metrics can you measure for recommendations?

Why is diversity needed in recommendations?

How can you diversify a list of products?

How is a recommendation ranking pipeline structured?

Which candidate selection approaches do you know?

What loss functions would you use for the different approaches?

How would you approach a ranking problem from scratch?

In what order should the selected products be shown?

Are you familiar with modern recommendation approaches?

What is position bias?

What is a feedback loop?

How would we handle it if a million products were returned instead of 100?

What if the system returns a million candidate products rather than 100 or 1,000?

What algorithm do we need to retrieve a diverse set of items?

Which metrics do we need to measure?

Which is more suitable: Recall or Precision@K?