1 / 5
The first stage that narrows millions of items down to a few hundred is ___.
-
-
-
-
Candidate generation (retrieval) cheaply selects a shortlist; the expensive ranking model then orders only those candidates.
2 / 5
Training a model to order items by predicted relevance is called ___.
-
-
-
-
Learning to rank (LTR) optimises the ordering of results, using signals like clicks and dwell time as training labels.
3 / 5
A metric that rewards putting relevant items higher in the list is ___.
Normalised Discounted Cumulative Gain (NDCG) discounts relevance by position, so ranking good items near the top scores higher.
4 / 5
Inputs to the ranker like recency, popularity, and user affinity are ___.
-
-
-
-
Ranking features (signals) feed the model; tuning often means adding, weighting, or removing these to improve relevance.
5 / 5
Recommending mostly near-duplicate items hurts list ___.
-
-
-
-
Without a diversity term, ranking can collapse into very similar items; tuning balances relevance with diversity for a better experience.