Multi-view Ordinal Regression with Dimensionality Reduction

Ordinal regression (OR) builds up the classifier on the data that the different categories are in an order. The majority of the existing approaches consider OR as a single-view learning problem, and only a little work is presented for multi-view OR (MVOR). Moreover, they build up the model on the original data and do not take the high-dimensional data into account. In this paper, we propose an novel dimensionality reduction approach for multi-view ordinal regression (DMVOR), which has the ability to handle the MVOR problem with high-dimensional data. DMVOR performs dimensionality reduction on high-dimensional multi-view data by mapping it into subspaces and learns a classifier in each subspace. The view consensus constraint is proposed to ensure the label consistency between different views. Numerical studies have illustrated that DMVOR outperforms the compared OR algorithms.