Recommender Systems Quiz
Eight questions spanning the course — the recommendation problem, ranking metrics, matrix factorization, two-tower retrieval, session-based models, diversity re-ranking, and eCPM ad ranking.
Exercise
Which is an example of IMPLICIT feedback?
Need a hint?
Stated preference vs inferred behavior.
Exercise
What is the defining characteristic of collaborative filtering compared to content-based filtering?
Need a hint?
What data does it use — item features, or the interaction matrix?
Exercise
What does the logarithmic discount in NDCG accomplish?
Need a hint?
Compare a relevant item at rank 1 vs rank 10.
Exercise
When training matrix factorization on a sparse rating matrix, why must you fit only the observed entries rather than treating missing entries as 0?
Need a hint?
What does a missing rating actually mean?
Exercise
In a two-tower retrieval model, how is the candidate (item) tower used at serving time?
Need a hint?
When are item embeddings computed relative to the request?
Exercise
A six-month jazz listener opens the app and starts browsing workout playlists. Which model responds correctly, and why?
Need a hint?
What signal does classical CF use, and what does a session model use?
Exercise
Maximal Marginal Relevance re-ranks with score λ·relevance − (1−λ)·max-similarity-to-selected. What problem does the second term solve?
Need a hint?
What does a list of ten near-identical top-scoring items look like to the user?
Exercise
Ad A bids $2.00 with predicted CTR 1%. Ad B bids $0.50 with predicted CTR 6%. Ranked by eCPM, which wins and why?
Need a hint?
eCPM = CTR × bid × 1000.
Exercise
Ranking metrics like recall@k and NDCG@k decide whether a recommender is good. Which slot of the ML project loop do they define?
Need a hint?
They measure how good the ranked output is — they don't train the model.
Exercise
Transfer test — a technique this course did NOT teach: MixUp, which trains on convex combinations of pairs of examples and their labels. Which slot of the ML project loop does it modify?
Need a hint?
It creates new synthetic training examples.