An efficient method for learning nonlinear ranking SVM functions

The problem of learning ranking (or preference) functions has become important in recent years as various applications have been found in information retrieval. Among the rank learning methods, RankSVM has been favorably used in various applications, e.g., optimizing search engines and improving data retrieval quality. Fast learning methods for linear RankSVM (RankSVM with a linear kernel) have been extensively developed, whereas methods for nonlinear RankSVM (RankSVM with nonlinear kernels) are lacking. This paper proposes an efficient method for learning with nonlinear kernels, called Ranking Vector SVM (RV-SVM). RV-SVM utilizes training vectors rather than pairwise difference vectors to determine the support vectors, and is thus faster to train than conventional RankSVMs. Experimental comparisons with the state-of-the-art RankSVM implementation provided in SVM-light show that RV-SVM is substantially faster for nonlinear kernels, although our method is slower for linear kernels. RV-SVM also uses far fewer support vectors, and thus the trained models are much simpler than those built by RankSVMs while maintaining comparable accuracy. Our implementation of RV-SVM is accessible at http://dm.hwanjoyu.org/rv-svm.