Description:
Structured classification deals with a family of problems where a response variable possessing meaningful internal structure has to be predicted from a set of input variables. This includes prediction problems involving strings, trees, or graphs as well as collective classification problems, i.e. problems where multiple correlated outputs have to be jointly predicted. We propose a general framework that combines the effectiveness of discriminative methods such as Gaussian processes and Support Vector machines for learning classification functions with the elegance and benefits of probabilistic graphical models for representing structural dependencies. Our framework significantly extends the standard classification setting, leading to learning algorithms that have numerous interesting applications, ranging from information retrieval, natural language processing and speech recognition to computer vision and bioinformatics.
Joint work with Ioannis Tsochantaridis (Google), Yasemin Altun (Toyota Technical Institute), Thorsten Joachims (Cornell University) and Alex Smola
(NICTA)
Speaker(s):
Thomas Hofmann
|
