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。
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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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