[Adapt] Seminar topic: Graph-based Semi-Supervised Learning

[陶宇超]

Hi, Adapters:

This is an old topic starting from 2003, presented by 'Xiaojin Zhu' in his paper "Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions". Given a certain data manifold, with little labeled data, large amout of unlabeled data can be classified. It is useful when labeling data is time consuming and it is easy to find the graph structure among the data.

No neural network is used in this paper. This learning approach is based on the property of harmonic function. Analytic solutions can be solved for each unlabeled data. Also due to the extensibility of the graph, this approach can incorporate external classifiers, which can provide complementary information.

This approached has been applied on many NLP tasks like entity linking. Moreover, others change the analytic solution to an iterative solution, which is more efficient to solve.

To better understand this learning approach, I hope you can review Gaussian Field, basic matrix operations, harmonic functions, kNN and random walk. 

Here are some useful references:

Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions:
    https://vvvvw.aaai.org/Papers/ICML/2003/ICML03-118.pdf
Collective Tweet Wikification based on Semi-supervised Graph Regularization:
    http://www.aclweb.org/website/old_anthology/P/P14/P14-1036.pdf
Harmonic function:
    https://en.wikipedia.org/wiki/Harmonic_function
______________________________________________________________
陶宇超
Harry Tao 

上海交通大学 
Shanghai JiaoTong University 

电子信息与电气工程学院 
School of Electronic Information and Electrical Engineering 

手机/Mobile:86 18930665880 
邮箱/e-mail: harry.t.chao at gmail.com 
             flyinhigh at sjtu.edu.cn