SVMRank

SVMrank is an instance of SVMstruct for efficiently training Ranking SVMs as defined in [Joachims, 2002c]. SVMrank solves the same optimization problem as SVMlight with the '-z p' option, but it is much faster. On the LETOR 3.0 dataset it takes about a second to train on any of the folds and datasets. The algorithm for solving the quadratic program is a straightforward extension of the ROC-area optimization algorithm described in [Joachims, 2006] for multiple rankings using the one...