引導式測驗 · 中等 等級
機器學習測驗中的歷史偏見
將歷史偏差與抽樣和標籤問題區分開來,並研究過去的社會過程如何塑造模型目標和決策。
問題 1 的 8
A model accurately learns patterns in prior hiring decisions that excluded a group. Which source of bias is most directly involved?
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 model accurately learns patterns in prior hiring decisions that excluded a group. Which source of bias is most directly involved?
- Historical bias
- Random initialization
- Software version drift
- Data encryption failure
- Why may collecting more records from the same historical process fail to remove historical bias?
- Larger data always causes more errors
- The source process can remain unequal even when represented accurately
- The model cannot train on old records
- Sampling size determines legal status
- A health algorithm uses past spending to predict medical need. What risk did Obermeyer and colleagues identify?
- Costs are always unrelated to care
- Spending is a perfect measure of illness
- Past spending can reflect unequal access, understating need for some patients
- The algorithm used no health data
- Which question helps identify a potentially biased training target?
- How many GPUs were used?
- What color is the interface?
- Who wrote the first line of code?
- Does the label measure the intended goal or a past institutional decision?
- How does historical bias differ from representation bias?
- Historical bias reflects inequity in the process; representation bias concerns who is missing or under-sampled
- 它們是相同的術語
- Representation bias only occurs after deployment
- Historical bias requires missing rows
- Why may removing a protected attribute fail to resolve historical bias?
- The model cannot use other features
- Other variables can act as proxies for historically structured differences
- Removing it automatically improves fairness
- Labels become independent of history
- Which action addresses a flawed historical target most directly?
- Collect more of the same target only
- Randomize the user interface
- Reconsider the target and compare it with the legitimate outcome the system should support
- Increase model size
- Can a fairness metric alone determine whether a historical outcome is legitimate?
- Yes, any single metric settles it
- Only if the model is small
- Yes, if the dataset is public
- No; legitimacy requires context and judgment beyond a metric