問題文
A hiring-support model is found to score applicants from one region lower. An engineer proposes deleting the region column from the dataset and considers the matter closed. Why is that step insufficient on its own?
選択肢
- Deleting a column always reduces accuracy, and the resulting drop would be larger than the disparity the team was trying to remove in the first place.
- Region is not a protected attribute, so removing it has no effect at all, and the scores can be compared as they stand.
- The disparity is caused by the loss function rather than by the input columns, so no change to the dataset can affect it and only retraining with a new objective helps.
- Other columns correlate with region, so the model can still recover the same signal indirectly and the disparity can persist.