引導式測驗 · 中等 等級
Z 分數和 IQR 異常值檢測測驗
檢查 z 分數計算、IQR 柵欄、說明性標記值、偏差和樣本大小限制、調查、僅訓練閾值和漂移。
問題 1 的 8
How is a z-score calculated relative to a reference group?
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。
- How is a z-score calculated relative to a reference group?
- Subtract the reference mean from the value and divide by the reference standard deviation
- Subtract the median and divide by the interquartile range
- Divide the value by the number of observations only
- Subtract the value from the upper IQR fence
- For illustrative quartiles Q1 = 10 and Q3 = 18, what is the upper 1.5-IQR fence?
- 18, because Q3 is itself the upper fence
- 22, adding half the IQR
- 42, adding three IQRs to Q3
- 30, because IQR = 8 and 18 + 1.5 × 8 = 30
- Under that constructed upper fence of 30, which value is flagged on the high side?
- 29, because it is above Q3
- 18, because it equals Q3
- 32, because it exceeds the fence
- 10, because it equals Q1
- Why can a simple z-score rule mislead on a heavily skewed dataset?
- Skew changes a value's units from numbers to categories
- A z-score becomes undefined whenever a high value occurs
- The mean and standard deviation may poorly summarize a long, valid tail
- Quartiles and standard deviations are always identical under skew
- Which part of the data does the interquartile range span?
- The distance from the minimum to the maximum
- The interval containing exactly 95% of a normal distribution
- Only values above the mean
- The middle half, from the first to the third quartile
- A laboratory value exceeds both a z-score threshold and an IQR fence. What should happen next?
- Delete it automatically because two methods agree
- Investigate source records and clinical context before deciding whether it is erroneous
- Conclude the patient result is impossible
- Replace it with the dataset mean before checking the source
- When should an ML team fit an outlier threshold used during model evaluation?
- After inspecting every test label and selecting the best threshold
- On the training partition, then apply the defined rule to evaluation data
- On all future deployment data before training begins
- Separately on the test set so its own labels guide preprocessing
- A once-stable sensor process changes after a hardware upgrade. What should the team do with its old outlier rule?
- Recheck the reference population, thresholds and downstream effects
- Delete every newly flagged reading without review
- Keep it unchanged because historical thresholds are universally valid
- Assume any new flags prove the upgrade failed