Dimensionality Reduction for Ordinal Classification
Mouad Zine-El-Abidine, Helin Dutağacı, David Rousseau · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Many unsupervised and supervised dimension reduction techniques are available for visualization and interpretation of high-dimensional data for classification tasks. While the unsupervised techniques do not employ the class information at all, most supervised algorithms are blind to the order of classes in ordinal classification problems. In this paper, we propose a novel and intuitive dimension reduction technique specifically designed for visualization of high-dimensional features in ordinal classification tasks. The technique is an iterative process, where at each iteration a search is conducted in the high-dimensional space to find the viewpoint from which the centers of adjacent classes are seen most distant from each other. The data is then projected to the lower dimensional space defined by the optimum viewpoint. The iteration is terminated when the desired dimensionality is achieved. Experimental results on various ordinal datasets demonstrate that our technique can be used as a complementary tool to the classical dimensionality reduction methods.