RV-SVM: An Efficient Method for Learning Ranking SVM

Abstract

Learning ranking (or preference) functions has become an important data mining task in recent years, as various applications have been found in information retrieval. Among rank learning methods, ranking SVM has been favorably applied to various applications, e.g., optimizing search engines, improving data retrieval quality. In this paper, we first develop a 1-norm ranking SVM that is faster in testing than the standard ranking SVM, and propose Ranking Vector SVM (RV-SVM) that revises the 1-norm ranking SVM for faster training. The number of variables in the RV-SVM is significantly smaller, thus the RV-SVM trains much faster than the other ranking SVMs. We experimentally compared the RV-SVM with the state-of-the-art rank learning method provided in SVM-light. The RV-SVM uses much less support vectors and trains much faster for nonlinear kernels than the SVM-light. The accuracies of RV-SVM and SVM-light are comparable on relatively large data sets. Our implementation of RV-SVM is posted at http://iis.postech.ac.kr/tv-svm.