引導式測驗 · 中等 等級
多標籤分類測驗
涵蓋非獨佔標籤、二元相關性和分類器鏈、標籤冪集以及漢明損失如何對單一標籤錯誤進行評分。
問題 1 的 8
A news story can receive both “politics” and “finance” tags. Which target setup fits this task?
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 news story can receive both “politics” and “finance” tags. Which target setup fits this task?
- Multi-label classification only works with image data
- Multi-class classification always uses more labels than multi-label classification
- Multi-label classification allows zero, one, or many labels per example instead of exactly one
- There is no meaningful difference between the two
- When movie tags such as “thriller” and “action” often co-occur, what limitation does independent binary relevance have?
- It ignores potential correlations between labels
- It cannot be implemented with standard classifiers
- It requires labeling every possible combination of labels as its own class
- It only works for exactly two labels
- How do classifier chains address binary relevance's limitation, according to the guide?
- By ignoring label order entirely
- By training only a single classifier for all labels combined
- By removing labels that co-occur too often
- By feeding each classifier the predictions of prior classifiers in the chain as extra input
- What problem does the label powerset approach face as described in the guide?
- It cannot handle any label correlations
- It struggles when the number of possible label combinations grows too large relative to training data
- It requires a sigmoid activation function
- It only works with numeric input
- Why is standard accuracy considered a poor evaluation metric for multi-label classification, per the guide?
- Because accuracy cannot be computed for binary problems
- Because accuracy requires GPU acceleration
- Because a prediction can be partially correct, which plain accuracy does not capture well
- Because multi-label problems never have a correct answer
- For a neural network that may assign several independent tags, which output setup does the guide describe?
- Sigmoid applied independently to each output node
- ReLU applied to each label
- Tanh applied to a single combined output
- Softmax applied to pairs of labels
- In a story-tagging task where “politics” and “finance” can both apply, why is a standard softmax output a poor fit?
- It makes class probabilities compete and sum to one, although multiple tags may be true at once
- It gives each possible tag a separate binary output
- It cannot represent more than two labels
- It removes the need for training examples
- According to the futureOutlook section, which baseline approach remains common in production despite not modeling label correlations?
- Label powerset
- Binary relevance
- Classifier chains
- Graph-based label dependency modeling