self-learning

逐次迭代，
Repeat：

1. train model f^*from labelled set
2. Apply f^*to the unlabeled date set.（get Pseudo-label）
3. add a set of unlabeled data to labelled data。jump to 1.

Hard Label V.S. Soft Label

极大似然估计（Maxmize likelihood）

Entropy-based Regularization

没太听懂
Semi-SVM

Smoothness Assumption
x1与x2之间有high density region

example：

Cluster（簇） and then Label

Graph-based Approach

K Nearest Neighbor