多分类(Multi-Classification)
One-Versus-All (OVA) Decomposition
以逻辑回归为例,其思路是将其中一类和剩下的类分开,做二分类,并对全部类做次操作,这样便有了K个逻辑回归分类器,只要取其中概率最大hypothesis所对应的分类作为分类结果即可。
- for \(k \in \mathcal { Y }\) , obtain \(\mathbf{w}_{[k]}\) by running logistic regression on
- return \(g ( \mathbf { x } ) = \operatorname { argmax } _ { k \in \mathcal { Y } } \left( \mathbf { w } _ { [ k ] } ^ { T } \mathbf { x } \right)\)
其优缺点是:
- pros: efficient ,can be coupled with any logistic regression-like approaches
效率高,可以和类似逻辑回归的算法(输出概率的算法)结合 - cons: often unbalanced D[k] when K large
如果K太大会导致数据不平衡
One-Versus-One (OVO) Decomposition
其基本思路是将其中一类和剩下的类中的一类做二分类,然对全部分类器执行该操作(组合数就是分类器数),那么
- for \(( k , \ell ) \in \mathcal { Y } \times \mathcal { Y }\) , obtain \(\mathbf { w }_ { [ k , l ] }\) by running logistic regression on
- return \(g ( \mathbf { x } ) = \text { tournament champion } \left\{ \mathbf { w } _ { [ k , \ell ] } ^ { T } \mathbf { x } \right\}\)
其优缺点是:
- pros: efficient (‘smaller’ training problems), stable, can be coupled with any binary classification approaches
更有效率更加稳定,可以结合任何二分类方法 - cons: use \(O(K^2) \,\mathbf { w }_ { [ k , l ] }\),more space, slower prediction, more training。
需要训练\(O(K^2)\) 个 \(,\mathbf { w }_ { [ k , l ] }\),占用更多的时间和空间。