引導式測驗 · 中等 等級
熵與資訊增益測驗
檢查二元熵、加權子雜質、數值資訊增益範例、貪婪分割、過度擬合和度量限制。
問題 1 的 8
Using base-two logarithms, what is the entropy of a node with equally many yes and no labels?
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。
- Using base-two logarithms, what is the entropy of a node with equally many yes and no labels?
- A value determined only by the feature scale
- Zero bits because the classes cancel
- Two bits because there are two labels
- One bit
- A child node contains only one class. What is its label entropy?
- It equals the parent entropy
- Zero, because its class label is pure
- It grows with the feature's units
- One bit for every example in the node
- Why are child entropies weighted by child size when computing information gain?
- To give every child equal influence regardless of sample count
- To use a child's entropy itself as its numerical weight
- Each child should contribute in proportion to the examples it contains
- To sum child impurities without normalizing for parent size
- A balanced binary parent has one bit of entropy. Two equal-sized children each have about 0.811 bits. What is the split's gain?
- About 0.189 bits, the parent minus weighted children
- One bit, because the parent is balanced
- About 1.811 bits, the sum of parent and child entropy
- About 0.811 bits, the average child entropy
- At one node, what does an entropy-based classification tree seek among candidate splits?
- The feature with the largest raw numeric values
- A large local reduction in weighted label impurity
- The split with the most children regardless of labels
- The feature with the strongest causal effect
- Why does the largest training-set information gain at a node not guarantee the best deployed tree?
- The choice is local and may overfit small training groups
- Information gain is always negative for useful splits
- All candidate features necessarily yield identical gain
- A high-gain split proves future outcomes are deterministic
- A split sharply separates historical labels using a location proxy. What does high information gain fail to establish?
- That the training labels were divided into children
- That the children have a weighted entropy
- That the split reduced the observed label mixture
- That the feature is causal, appropriate, or fair to use
- If a team switches from base-two logs to another log base for every candidate at the same node, what happens?
- The training labels change
- Only pure-node entropy becomes positive
- Gain values rescale, but the candidate ranking stays the same
- Information gain stops requiring child-size weights