引導式測驗 · 中等 等級

基於對話階段的推薦系統測驗

測試基於會話的建議:有序互動、會話邊界、序列模型、下一個評估、基線、隱私和短期

問題 1 的 8

A visitor has no account history but has viewed several products in one visit. Which information can a session recommender use?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. A visitor has no account history but has viewed several products in one visit. Which information can a session recommender use?
    • A complete long-term profile that does not exist
    • The ordered interactions from the current visit
    • Only the product catalog’s average price
    • The visitor’s identity inferred from device ownership
  2. Why should a short sequence not automatically become a durable preference profile?
    • Session data always reveal stable taste
    • A session contains no item information
    • Long-term profiles cannot be corrected
    • It may reflect a temporary task, accidental clicks, or incomplete evidence
  3. How does a recurrent session model use item order?
    • It updates a hidden representation as events arrive in sequence
    • It treats every interaction as an unordered count
    • It uses only the first item viewed
    • It predicts without receiving any events
  4. Why test on later sessions instead of randomly splitting interaction rows?
    • Random splitting guarantees future conditions
    • Later sessions contain no user behavior
    • A chronological or session-level split reduces leakage between training and evaluation
    • Chronology matters only for image models
  5. What does Recall@K measure in a next-item task?
    • Whether a user liked every item in the catalog
    • Whether the observed next item appears within the top K recommendations
    • The time required to train the model
    • Whether the session boundary was legally correct
  6. Why can treating every unclicked item as disliked distort training data?
    • The user may never have noticed or been shown the item
    • Clicks are always accidental
    • Unclicked items are identical to purchases
    • A model cannot use negative examples
  7. What additional behavior should an online evaluation monitor beyond clicks?
    • 僅使用模型參數數量
    • Only page-load color
    • The number of candidate items before ranking
    • Returns, cancellations, coverage, and user controls
  8. RepeatNet-style models distinguish repeat behavior from exploration in which way?
    • Both always recommend the most popular item
    • Repeat behavior means repeating the same search query
    • The system models both returning to an earlier item and suggesting a different one
    • Exploration prevents a user from seeing prior items