Measure the impurity drop before and after a binary split using the Gini index or entropy, to compare candidate features when building a decision tree.
决策树分裂选择增益最大的特征。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Entropy | entropy is the negative sum of class share times log base two share | Information theory |
| Information gain | gain = parent entropy minus the weighted child entropies | Quinlan ID3 |
| Gain ratio | gain ratio divides by intrinsic value to curb multi-valued bias | Quinlan C4.5 |
信息熵 H=−Σpᵢlog₂pᵢ;基尼不纯度 G=1−Σpᵢ²,均衡量节点混杂度。