机器学习的类型
根据输出空间种类进行分类
1. 二分类,输出 Y = {−1, +1}:
PLA是线性二分类,二分类还有其他的
2. 多分类,输出 Y = {1, 2, · · · , K} (abstractly) 二分类是多分类的特殊情况( K = 2)
Coin Recognition Problem
3. 回归,输出 Y = R or Y = [lower, upper] ⊂ R (bounded regression)
4. Structured Learning: Sequence Tagging Problem
sentence ⇒ structure (class of each word) 输出 Y = {PVN, PVP, NVN, PV, · · · }, not including VVVVV
总结,我们可以根据输出空间的类型来对学习算法进行分类
根据标签进行分类
1.监督学习 supervised learning: every xn comes with corresponding yn
2. 无监督学习 Unsupervised: every xn comes without corresponding yn
3. 半监督学习 Semi-supervised: some xn comes with corresponding yn (标签获取困难或者费时费力时 采用这种方法)
4. 强化学习 Reinforcement Learning
总结,我们可以根据不同的标签来对学习算法进行分类
根据数据的训练方式对算法进行分类
1 batch learning 从所有已知的数据中训练学习
2 online learning 每次训练一个数据,然后根据结果对模型进行修正,然后训练下一个数据
3 activate learning Learning by ‘Asking’
根据输入的特征类型进行分类
1. Concrete Features
通常这些特征来自一定的先验知识,比如(size, mass) for coin classification, 并且具有sophisticated physical meaning 。
2. Raw Features
这些特征通常具有‘simple physical meaning, 比如图像的像素 。因为数据简单抽象,机器训练起来比较难
raw features often need human(常常称之为特征工程) or machines to convert to concrete ones
3 Abstract Features
‘no physical meaning’; thus even more difficult for ML; again need ‘feature conversion/extraction/construction
总结
参考资料: