Quiz dotyczący modeli rekomendacji dwóch wież

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. What does each tower produce in a standard two-tower recommender?
   • A vector representation of the query/user or candidate item
   • A shared label for all users and items
   • A direct causal effect estimate
   • An ordered list of every catalog item
2. Why can item embeddings be precomputed?
   • Every user has identical preferences
   • The item tower can encode candidates independently of the current user
   • The model has no item features
   • The item labels are known in advance
3. Which operation commonly retrieves candidates from embeddings?
   • A dendrogram cut
   • A per-item full cross-encoder over the entire catalog
   • Nearest-neighbor or maximum-inner-product search
   • A residual Q-Q plot
4. Why is candidate recall important before ranking?
   • It makes the catalog finite
   • High recall guarantees calibrated probabilities
   • It replaces all feature preprocessing
   • A ranker cannot choose an item absent from the retrieved set
5. Which interaction is difficult for independently encoded towers to represent directly?
   • They cannot use user features
   • They cannot directly compute arbitrary pair-specific features at retrieval time
   • They require every item to have a label
   • They only support one candidate
6. What can sampled negatives affect during training and evaluation?
   • The item catalog's release date
   • The dimensionality of user vectors only
   • The learned discrimination task and how realistic offline metrics are
   • The number of towers
7. What can a high dot-product score be safely interpreted as?
   • Proof the item was eligible
   • A calibrated probability of satisfaction
   • A causal effect of showing the item
   • A model similarity score for retrieval
8. Why version item index and tower preprocessing together?
   • Mismatched embeddings or transforms can make similarity comparisons invalid
   • The index otherwise becomes a supervised target
   • It sets negative samples to zero
   • It changes the model into a single tower