引導式測驗 · 中等 等級
評估推薦系統:NDCG、命中率和更多測驗
測試前 K 排名指標定義、每個指標捕獲的內容、標籤如何影響解釋,以及為什麼離線分數需要即時結果驗證。
問題 1 的 8
A list has 10 recommendations, of which 4 are relevant under the evaluation labels. What is Precision@10?
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。
- A list has 10 recommendations, of which 4 are relevant under the evaluation labels. What is Precision@10?
- 0.4, because 4 of the first 10 items are relevant
- 0.6, because 6 items are irrelevant
- 4, because precision is the relevant-item count
- 0.1, because the cutoff is 10
- Recall@K 在推薦評估中測量什麼?
- Whether at least one item in the list was clicked
- The average position of every catalog item
- The share of all labeled relevant items that appears in the top K
- The share of the top K list that is relevant
- 在常見的 Hit Rate@K 慣例下,何時將使用者視為命中?
- Every item in the top-K list is relevant
- The user clicks every recommendation
- The first result has a higher score than the second
- At least one held-out relevant item appears in the user’s top-K list
- NDCG 相較於無序相關推薦計數增加了什麼?
- It computes whether any one item appears anywhere in the catalog
- It accounts for ranking position by discounting lower-ranked relevant results and normalizing by an ideal ranking
- It measures only catalog size and item coverage
- It replaces relevance labels with click-through revenue
- 為什麼報告應說明其相關性標籤與指標慣例?
- All relevance labels are interchangeable once K is selected
- A single standard convention is mandated across every recommender
- Metric labels do not affect reported results
- Clicks, purchases, and ratings represent different behaviors, and metric implementations can vary
- An offline NDCG score improves. What can the team safely conclude?
- The model ranked the chosen evaluation labels better under that offline setup; live impact still needs validation
- The product will necessarily earn more revenue
- Users will definitely be more satisfied
- The model is ready to launch without guardrails
- 為什麼歷史點擊紀錄可能是偏頗的相關來源?
- Historical logs contain no ranking information
- User activity can only bias Recall, never Precision or NDCG
- The previous ranker influenced which items users had a chance to see
- Clicks always reveal objective relevance without exposure effects
- Which evaluation plan is most informative before a major recommender launch?
- Compare global click totals without checking user-level exposure
- Report offline metrics with definitions and guardrails, then test live outcomes where suitable
- Choose the model with the highest Hit Rate and skip product metrics
- Optimize a single offline score until it reaches 1.0