問題文
A dataset of product photographs was collected by scraping a catalog. Many products appear several times with small differences: a different crop, a watermark, or a slightly different white balance. The team split the data randomly. Validation accuracy is far above what live traffic shows. What is the most likely cause?
選択肢
- The validation set is too small, so its estimate has high variance and the situation will correct itself once more photographs are labeled.
- Near-duplicate photographs of the same product ended up on both sides of the split, so the validation set measures recall of seen items rather than generalization.
- The live traffic contains a different set of products, so the gap reflects a change in the label space rather than a flaw in the split.
- The watermarks act as noise and lower the live accuracy.