Ordinal multi-class Molecular Classification

Abstract

When a cancer grows, it progresses from one stage to another, which can been seen as a sequence of ordered phases. Current research on multi-class molecular classification typically treats the classes on a nominal scale and thus does not take any relation between classes into account. The ordering is however valuable information which may be used to improve the predictive power of a classifier. A few ordinal classifiers have been published, but they have not been applied in the analysis of molecular data, where there are only a limited number of samples in comparison to the number of features. This paper describes a comparative study in which current ordinal classifiers are benchmarked in a molecular analysis of gene expression. This helps to determine whether using the relation between classes can help to improve the prediction results. The results of the comparison study shows that there is not a lot of difference in performance between nominal and ordinal classifiers evaluated on real datasets. Several experiments were executed to further investigate any difference between both types of classifiers. It seems that by selecting monotonous features (i.e. features that correlate linearly with the class labels), the performance of nominal classifiers can be significantly improved. This allows for the usage of well-known and less complex classifiers, which is beneficial in p > n problems.