Generalization Bounds for Ordinal Regression Algorithms via Strong and Weak Stability.

Abstract: Ordinal regression shares properties of both classification and regression, since the goal function maps the instance space into the integer set. In this paper, we studied generalization properties of ordinal regression algorithms, and we focused on the performance of goal function effecting on individual examples. The strong (weak) loss stabilities and strong (weak) score stabilities for standard regression algorithms we defined only consider changing one element of training set. With the definitions of stabilities for ordinal regression algorithms and the clipped loss function, four kinds of generalization bounds are given in this paper. [Copyright &y& Elsevier]