Q统计量
Q统计量是衡量Ensemble模型中两两分类器（pair wise）有多不同的一个指标。
比如两个分类器分类结果
| $D_k correct(1)$ | $D_k wrong(0)$ | |
|---|---|---|
| $D_i correct(1)$ | a | b |
| $D_i wrong(0)$ | c | d |
其中归一化（除以总样本）之后$a+b+c+d=1$，那么Q统计量为
Intuition
For statistically independent classifiers, $Q_{i,k} = 0$. $Q$ varies between -1 and 1. Classifiers that tend to recognize the same objects correctly will have positive values of $Q$, and those which commit errors on different objects will render $Q$ negative. For a set of $L$ classifiers, the averaged $Q$ statistics of all pairs is taken.
Refs
- Ten Measures Of Diversity In Classifier Ensembles: Limits For Two Classifiers
- Diversity in Classifier Ensembles
