Support Vector Machines for Ranking Learning: The Full and the Truncated Fixed Margin Strategies

Two known SVM-based approaches to ranking learning (ordinal regression estimation, supervised pattern recognition with ordered classes) are scrutinized as different generalizations of the classical principle of finding the optimal discriminant hyperplane in a linear space. Easily verifiable natural conditions are found under which the training result obtained by the computationally much more attractive truncated technique is completely equivalent to the hypothetical strict solution. The numerical procedures are essentially simplified for both techniques.