Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo Cy-Umunara Icyitegererezo Cyikibazo

Kurikirana umukoresha nibintu bitandukanye kodegisi ukoresheje uburyo bworoshye bwo kugarura ibintu, hanyuma umenye ibicuruzwa biva hamwe nabakandida bibutse hamwe nu rutonde rwo hasi.

Inzira zijyanye no kuyoboraUburyo bubiri bwo gusaba umunara
Ikibazo 1 Bya 8

What does each tower produce in a standard two-tower recommender?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  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