引導式測驗 · 中等 等級

聚合偏差解釋測驗

區分聚合偏差和代表性差距,並評估合併模型的共享輸入輸出假設何時在子組中失敗。

相關引導路徑AI 模型中的聚合偏差
問題 1 的 8

A pooled model fits the average population but performs poorly because subgroup relationships differ. Which bias is this?

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 pooled model fits the average population but performs poorly because subgroup relationships differ. Which bias is this?
    • Aggregation bias
    • Only random measurement error
    • Encryption bias
    • A user-interface defect
  2. How does aggregation bias differ from representation bias?
    • Aggregation bias concerns a shared model that does not fit distinct patterns; representation bias concerns coverage gaps
    • They mean exactly the same thing
    • Representation bias is only about labels
    • Aggregation bias occurs only after launch
  3. Which scenario illustrates possible aggregation bias?
    • A model has no labels
    • One risk function is used for children and adults despite differing clinical patterns
    • A dataset has duplicate rows
    • A server times out
  4. Why can a strong overall score fail to reveal aggregation bias?
    • It is not a real metric
    • It automatically averages away every class
    • Pooled performance can conceal subgroup errors or miscalibration
    • It measures only training speed
  5. Which evaluation practice helps detect an unsuitable pooled model?
    • Report only one overall accuracy score
    • Examine relevant errors and calibration by subgroup with uncertainty
    • Use the training set as the test set
    • Remove all context variables
  6. Which is a possible response when evidence shows groups have different relationships?
    • Consider conditional, hierarchical, or group-specific modeling and evaluate tradeoffs
    • Always use separate models for every person
    • Ignore the pattern if average accuracy is high
    • Delete all subgroup data
  7. Why should teams avoid automatically creating separate thresholds for each group?
    • Thresholds cannot affect decisions
    • Separate thresholds are always illegal
    • They can have sparse-data, operational, ethical, and legal tradeoffs
    • They make subgroup evaluation impossible
  8. What supports choosing a pooled model?
    • It is easiest to deploy
    • Evidence that its shared assumptions fit the intended groups and purpose
    • A vendor’s claim alone
    • A single training score