Maximizing upgrading and downgrading margins for ordinal regression (Q1935916)

The paper is a contribution to the theory of classifying a set of objects into a set of ranked classes. Classifying an object in a class with a higher rank than its actual rank is called upgrading error. Similarly classifying an object in a class of lower rank is called downgrading error. Upgrading and downgrading errors are considered as two different classifier quality criteria to be taken into account. A support vector machines approach to maximization of undergrading and downgrading margins as an effective method of minimizing classification errors is proposed. A biobjective optimization problem is used to maximize simultaneously the upgrading and downgrading margins. It is shown that the resulting set of Pareto-optimal solutions can be obtained by solving one quadratic optimization problem. Several illustrative examples in the concluding part of the paper help to understand the proposed approach and applicability of the results.