A Support Vector Method for Multivariate （多元支持向量方法）.pdf

A Support Vector Method for Multivariate Performance Measures Thorsten Joachims tj@ Cornell University, Dept. of Computer Science, 4153 Upson Hall, Ithaca, NY 14853 USA Abstract timizing many different variants of convenient and tractable performance measures, aiming to find one This paper presents a Support Vector that performs well for the application specific per- Method for optimizing multivariate non- formance measure after post-processing the resulting linear performance measures like the F - 1 model (e.g. (Lewis, 2001; Yang, 2001; Abe et al., 2004; score. Taking a multivariate prediction ap- Caruana Niculescu-Mizil, 2004)). However, in par- proach, we give an algorithm with which such ticular for non-linear performance measures like F - 1 multivariate SVMs can be trained in poly- score or PRBEP, the relationship to tractable mea- nomial time for large classes of potentially sures is at best approximate and requires extensive non-linear performance measures,