Ordinal Regression as Multiclass Classification

Recently, an interesting framework was proposed to reduce the ordinal regression to the binary classification [1]. It made two independent assumptions: target functions are rankmonotonic; or rows of loss matrix are convex. Both of the two assumptions impose some restrictions on its application, because they may not be reliable in practice. This paper presents a novel reduction framework free of such restrictions, in this sense, which is more efficient and versatile in real world applications. Experiments on several datasets empirically validate its effectiveness. The contribution of our work is that it proves the fact that the ordinal regression is equivalent to the regular multiclass classification whose distribution is changed.