An improved ordinal regression approach with Sum-of-Margin principle

In this paper, we propose a new support vector approach for ordinal regression, which maximizes the sum of the margins of parallel discriminant hyperplanes. For ordinal regression, there are two strategies to take on the large margin principle: the fixed margin principle and the sum-of-margin principle. While the fixed margin strategy requires that the margins between two neighboring classes are equal and fails to define the thresholds of different ranks uniquely and directly, the Sum-of-Margin strategy is to maximize the sum of margins and the threshold defining each rank is unique and can be obtained directly. However, the performance of the traditional support vector ordinal regression method based on the Sum-of-Margin Principle is unsatisfactory because of unreasonable definition of empirical errors of training data. To solve this problem, we use different constraints and a new support vector ordinal regression algorithm is developed. The experiment results verify the effectiveness and efficiency of the proposed approach.