引導式測驗 · 中等 等級
「一對其餘」與「一對一」策略測驗
測試 OvR 和 OvO 分類器計數、分數比較和投票,以及成對模型何時適合具有昂貴擬合的核方法。
問題 1 的 8
For a classification problem with 6 classes, how many binary classifiers does the one-vs-rest strategy train?
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。
- For a classification problem with 6 classes, how many binary classifiers does the one-vs-rest strategy train?
- 15
- 6
- 30
- 36
- For the same 6-class problem, how many classifiers does one-vs-one train?
- 6
- 12
- 21
- 15
- In one-vs-rest, how is the final predicted class chosen among the K trained classifiers?
- The classifier with the highest confidence score
- Majority vote among all classifiers
- The first classifier that returns positive
- Random selection among positives
- In one-vs-one, how is the final predicted class determined?
- The pairwise classifier trained on the most data wins
- Only the first pairwise classifier's output is used
- Each pairwise classifier casts one vote and the class with the most votes wins
- Classes are ranked by training set size
- Why is OvO traditionally the default multiclass strategy for kernel support vector machines?
- Because it always gives higher accuracy than OvR
- Because each pairwise classifier trains on a smaller subset of data, offsetting having more classifiers
- Because it requires no probability calibration ever
- Because OvR cannot be implemented mathematically for SVMs
- What problem can arise specifically in OvR when comparing scores across the K independently trained classifiers?
- The scores are always identical across classifiers
- The classifiers cannot output any confidence measure
- OvR cannot produce ties
- The raw scores may not be on a comparable scale since each classifier was trained separately
- Which algorithms are described as natively handling multiple classes without needing OvR or OvO wrappers?
- Neural networks and gradient-boosted trees using softmax or multinomial objectives
- Basic support vector machines
- Standard logistic regression solvers restricted to binary output
- None of the algorithms mentioned handle multiclass natively
- A rare category is modeled against all other examples in OvR. What training-data pattern may make that classifier difficult to fit?
- Each classifier sees no negative examples
- Each classifier requires a separate test set
- Each classifier's training set can be highly imbalanced since it pits one class against all others combined
- OvR classifiers cannot be trained in parallel