Transductive Experimental Design (TED)

 

 

·      Short Introduction: TED is a very simple and effective algorithm to select the most informative experiments x to get measurements y for learning a regression model y = f(x). We demonstrated in the special case of square error loss, the active learning is independent to measurements (labels). This allows us to design a very simple active learning algorithm (see the Algorithm 1 in our ICML paper) that avoids expensive retraining and can be implemented by only several lines of matlab codes. Though TED is designed for least-squares regression, it has shown very good results for also classification tasks, for example, text categorization.

 

 

 ·      Active Learning via Transductive Experimental Design 

Kai Yu, Jinbo Bi, and Volker Tresp

Proceedings of the 23rd  International Conference on Machine Learning (ICML 2006), July, 2006 

[pdf] [slides] [bibtex] 

 

·      Matlab Code and Data: actvted_demo.zip (2.2 MB, including a toy problem and an experiment on a text corpus benchmark),   readme.txt