Learning from small data: A pairwise approach for ordinal regression

Ordinal regression which aims to classify instances into ordinal categories has numerous applications. As a supervised learning problem, a large number of labeled data is needed to train an accurate model, in particular when the number of categories is large. Learning an effective ordinal classifier from a small dataset is a challenging task. This paper proposes a framework to transform the ordinal regression problem to a binary classification problem and then recover the ordinal information from the binary outputs. The labeled instances are paired up to train a binary classifier, and therefore, the number of training points is squared, which alleviates the lack of training points. The transformed binary classification problem is solved by a pairwise SVM method. Experimental results demonstrate that on 12 widely used benchmarks, the proposed method is effective comparing with the state-of-the-art ordinal regression methods.