半监督学习

Algorithm Details : 
Label Propagation Semi Supervised Learning : 
Label Propagation：原理， 利用calculate similarity by e.g. the rbf kernel) 进行传播
1) Actually LP (Label Propagation) is Transductive [2] setting of Semi-supervised Learning. 
2) Transductive means the unalabeled examples it self is test data. There is no outo of sample data point. eg: n=l+u (l examples will have label associated , u no of examples we have predict the labels).

3)  One way to predict the labels for u examples is LP algorithms, it has been proposed by Zhu and Gharamani ("LP-ZGL  Algo " famously known) in 2002.

4) It's very simple algorithm to explain. 

 Step1: Construct the graph of n data points, so input is n*d (no examples and dimension) output is Transistion Proability Matrix (T).
Step2 : Prepare label Vector (Y0) as for l exmples is either +1 or -1 and u examles it's all zeros. 
Step3:  Multiply T*Y0 to get Y1  (in Y1 some of examples  will get value from 0 to non-zero value)。T是传递矩阵
Step4: Repeat Step3 till it convereges (i.e. , all the u examples attain a value from zero (0) to either +1 or -1)

n = l + u ，l是已有的label， u是未标注的样本数据

里边关键的一个处理是，已标注的节点的label在传递过程中的处理，即传递过程中已有的label变不变的问题。