03 Types of Learning
从四个维度介绍常见机器学习类型，见详细课件。
Output Space
从$Y$的维度考虑，不同的输出空间，对应不同的机器学习算法。
Binary Classification
二分类问题，输出空间为$Y={−1, +1}$。常见例子比如：
- credit approve/disapprove
- email spam/non-spam
- patient sick/not sick
- ad profitable/not profitable
是极其重要的一类问题：
Core and important problem with many tools as building block of other tools.
Multiclass Classification
多分类问题，输出空间为$Y={1,2,\dots, K}$，二分类是$K=2$时候的特例。常见例子比如：
- coin recognition
- written digits ⇒ 0, 1, · · · , 9
- pictures ⇒ apple, orange, strawberry
- emails ⇒ spam, primary, social, promotion, update
Regression
回归问题，输出空间$Y=R$或者$Y=[lower, upper] \in R$，对应bounded regression。常见的例子比如：
- patient features ⇒ how many days before recovery
- company data ⇒ stock price
- climate data ⇒ temperature
统计学中被广泛研究：
Also core and important with many ‘statistical’ tools as building block of other tools.
Structured Learning
结构化学习，常见例子比如：
- sentence ⇒ structure (class of each word)(序列标注)
- protein data ⇒ protein folding
- speech data ⇒ speech parse tree
Huge multiclass classification problem (structure = hyperclass) without ‘explicit’ class definition.
Data Label
从data label $y_n$的有无、多少、形式划分：
- supervised: all $y_n$
- unsupervised: no $y_n$
- semi-supervised: some $y_n$
- reinforcement: implicit $y_n$ by goodness
Supervised Learning
Supervised learning: every $x_n$ comes with corresponding $y_n$.
比如二分类、多分类问题，都是典型的监督学习。
Unsupervised Learning
Unsupervised learning: diverse, with possibly very different performance goals.
无监督学习形式也很丰富，常见的比如：
- clustering
- ${x_n} ⇒ cluster(x)$
- unsupervised multiclass classification
- i.e. articles ⇒ topics
- density estimation
- ${x_n} ⇒ density(x)$
- unsupervised bounded regression
- traffic reports with location ⇒ dangerous areas
- outlier detection
- ${x_n} ⇒ unusual(x)$
- extreme ‘unsupervised binary classification’
- i.e. Internet logs ⇒ intrusion alert
Semi-supervised Learning
Semi-supervised learning: leverage unlabeled data to avoid ‘expensive’ labeling.
常见的比如：
- face images with a few labeled ⇒ face identifier (Facebook)
- medicine data with a few labeled ⇒ medicine effect predictor
详细解释见Semi-supervised learning。
Reinforcement Learning
Reinforcement: learn with ‘partial/implicit information’ (often sequentially).
样本形式$(x_n, y_n, goodness)$常见的比如：
- (customer, ad choice, ad click earning) ⇒ ad system
- (cards, strategy, winning amount) ⇒ black jack agent
Different Protocol
不同Protocol对应不同Learning Philosophy：
- batch: duck feeding
- online: passive sequential
- active: question asking (sequentially)(query the $y_n$ of the chosen $x_n$)
对应的训练数据也不相同：
- batch: all known data
- online: sequential (passive) data
- active: strategically-observed data
Batch Learning
一次性从所有已知数据中学习。
Batch supervised multiclass classification: learn from all known data.
- batch of (email, spam?) ⇒ spam filter
- batch of (patient, cancer) ⇒ cancer classifier
- batch of patient data ⇒ group of patients
Online Learning
序列地接受数据，然后更新模型。
Online: hypothesis ‘improves’ through receiving data instances sequentially
比如online spam filter, which sequentially:
- observe an email $x_t$
- predict spam status with current $g_t(x_t)$
- receive ‘desired label’ $y_t$ from user, and then update $g_t$ with $(x_t, y_t)$
PLA can be easily adapted to online protocol.
Active Learning
当模型没有把握的时候，把问题交给用户，从而获取高质量样本。
Active: improve hypothesis with fewer labels (hopefully) by asking questions strategically
Different Input Space
根据输入空间的含义划分。
Concrete Features
Concrete features: each dimension of $X \in R^d$ represents ‘sophisticated physical meaning’.
常见的比如：
- (size, mass) for coin classification
- customer info for credit approval
- patient info for cancer diagnosis
- often including human intelligence on the task
这些具体特征，有明确的含义，可解释性很强，同时easy for ML。
Raw Features
Raw features: often need human or machines to convert to concrete ones.
比如image pixels, speech signal等场景。
Abstract Features
Abstract: again need feature conversion/extraction/construction.
比如一些ID特征：
- student ID in online tutoring system (KDDCup 2010)
- advertisement ID in online ad system
