Relative margin induced support vector ordinal regression

As a classical ordinal regression model, support vector ordinal regression (SVOR) finds (r-1) parallel discriminant hyperplanes via maximizing the minimal margins between different ranks. The ordinal relation is guaranteed by explicit or implicit constraints. However, the minimal margin between adjacent ranks is only determined by minor patterns near the margin hyperplanes and others have no influence on the discriminant hyperplane learning. In order to reflect the contributions of these patterns, this paper proposes relative margin induced support vector ordinal regression (RMSVOR) models, in which the margin between a pattern and a discriminant hyperplane is depicted by a function of relative margin information to reflect its contribution on this hyperplane. The relative margin information is estimated by nearest neighbor chain to reflect the prior knowledge of the pattern in training set. The experiments, performed on discretized regression datasets and real ordinal regression datasets, demonstrate that RMSVOR is superior to previous ordinal regression models (SVOR, NPSVOR and NPHORM) and canonical multi-class classification models (OvASVM and OvOSVM).