引導式測驗 · 中等 等級

公平性不可能定理測驗

測試三分標準結果、基本率和預測假設、校準和錯誤率平衡的定義以及定理的限制。

相關引導路徑公平性不可能定理
問題 1 的 8

Which three conditions are central to the Kleinberg–Mullainathan–Raghavan result?

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. Which three conditions are central to the Kleinberg–Mullainathan–Raghavan result?
    • Accuracy, speed, and model size
    • Privacy, transparency, and consent
    • Calibration, balance for the positive class, and balance for the negative class
    • Precision, recall, and training time
  2. Under what common special case can the three conditions avoid the stated conflict?
    • Equal base rates across groups or perfect prediction
    • Different data formats
    • A larger training set alone
    • Any use of a human reviewer
  3. What does calibration within groups mean?
    • False-positive rates are identical
    • Every group receives the same number of positive decisions
    • No protected attribute is used
    • People with a given score have the corresponding observed outcome frequency within each group
  4. For a thresholded classifier, which rates does Chouldechova’s related error-rate-balance criterion compare?
    • Positive decision rates only
    • False-positive and false-negative rates across groups
    • Training and inference time
    • Data retention and deletion rates
  5. What role do differing base rates play in the incompatibility results?
    • With differing outcome prevalence and imperfect predictions, some criteria cannot generally all hold exactly
    • They guarantee all fairness criteria can hold
    • They matter only for privacy laws
    • They mean the model is necessarily inaccurate
  6. Does the theorem prove that fairness is impossible in general?
    • Yes, every fair system is mathematically impossible
    • Yes, it proves discrimination is lawful
    • No; it concerns specified statistical criteria under specified assumptions
    • No, because it only concerns privacy
  7. Does the result choose which fairness metric a policy maker should prioritize?
    • Yes, it always mandates calibration
    • No; it shows a conflict but leaves the value judgment and harm trade-off to decision makers
    • Yes, it always mandates equal error rates
    • It makes all metrics irrelevant
  8. Why must an auditor check how the outcome label was produced?
    • Fairness metrics do not use outcomes
    • Labels determine whether code compiles
    • The model only works with image labels
    • The theorem assumes a defined target, but labels may be incomplete or reflect structural disadvantage