Novel fuzzy rank discrimination measures for monotonic ordinal feature selection

Classification with ordinal monotonic constraints refers to the sample categories are preference-ordered, and monotonicity constraints between the sample categories and class labels needs to be guaranteed in the learned classification model. Feature selection aims at finding highly discriminative features to avoid degradation of learning performance, which has increasingly gathered attention in the field of monotonic classification. In this paper, we propose novel fuzzy rank discrimination measures to evaluate monotonic consistency for feature selection in ordinal dataset. The proposed measure fulfill the monotonicity property in the sense that value of measure does not decrease when a new candidate feature is added to the selected feature subset, which guarantees the optimal feature subset can be achieved without evaluating all possible combinations of features explicitly. A wrapper approach based on the feature ranking in terms of minimal-redundancy and maximal-relevance criterion is implemented to evaluate the performance of the proposed method for monotonic feature selection. The superiority of the proposed fuzzy rank discrimination measure over the existing measure, in terms of robustness and classification accuracy, is established extensively on several real-life monotonic datasets by using four monotonic classifiers.