Efficient SVM for Ranking

Ranking Vector Machine improves the efficiency of the standard Ranking SVM (as implemented in SVM-light) by (1) using L_1 norm instead of L_2 norm for regularisation; and (2) using a subset of the original instance vectors rather than the pairwise difference vectors as support vectors. It is said to be much faster than Ranking SVM if non-linear kernels are employed, though with a lower accuracy (especially on small datasets).